[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip / qa] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/mlp/ - Pony


Thread archived.
You cannot reply anymore.




File: New OP.png (1.54 MB, 2119x1500)
1.54 MB
1.54 MB PNG
Welcome to the Pony Voice Preservation Project!
youtu.be/730zGRwbQuE

The Pony Preservation Project is a collaborative effort by /mlp/ to build and curate pony datasets for as many applications in AI as possible.

Technology has progressed such that a trained neural network can generate convincing voice clips for any person or character using clean audio recordings as a reference. As you can surely imagine, the ability to create audio in the voices of any pony you like has endless applications for pony content creation.

AI is incredibly versatile, basically anything that can be boiled down to a simple dataset can be used for training to create more of it. AI-generated images, fanfics, wAIfu chatbots and even animation are possible, and are being worked on here.

Any anon is free to join, and there are many active tasks that would suit any level of technical expertise. If you’re interested in helping out, take a look at the quick start guide linked below and ask in the thread for any further detail you need.

EQG and G5 are not welcome.

>Quick start guide:
derpy.me/FDnSk
Introduction to the PPP, links to text-to-speech tools, and how (You) can help with active tasks.

>The main Doc:
docs.google.com/document/d/1xe1Clvdg6EFFDtIkkFwT-NPLRDPvkV4G675SUKjxVRU/edit
An in-depth repository of tutorials, resources and archives.
(Currently down, working on replacement)

>Active tasks:
Research into animation AI
Research into pony image generation
Preparations for /mlp/con PPP panel

>Latest developments:
GDrive clone of Master File now available (>>37159549)
SortAnon releases script to run TalkNet on Windows (>>37299594)
TalkNet training script (>>37374942)
GPT-J downloadable model (>>37646318)
FiMmicroSoL model (>>38027533)
FiMfic dataset (>>38029649)
Delta releases new GPT-J notebook + tutorial (>>38018428)
Another FIMfic downloader (>>38051136)
New DeltaVox (>>38064386)
New TTS Notebook: derpy.me/a2Ceh (>>38185345 dunky11)
New FiMfic GPT model (>>38308297 >>38347556 >>38301248 GothicAnon)
Work on projects/con panel (>>38714338)
Refined FimFic dataset (>>38391839 GothicAnon)
KoboldAi training tutorial (>>38422872 GothicAnon)
AI-Generated Song (>>38663340)
DALL-E image generation (38713667)
Latest Synthbot progress report (>>38392162 >>38397538 >>38410311 >>38440467 >>38486788)
Latest Clipper progress report (>>38714338)

PPP REDUB
-Ep1
youtu.be/gEXaFVw9J1o
derpy.me/ELksq

-Ep2
youtu.be/fIhj2bFYG4o
derpy.me/RHegy

-Ep 3
youtu.be/0BtJIrVhuuI

-Unused Clips
youtu.be/N2730oPqLzE
derpy.me/OKoqs

-Rewatch Premiere
derpy.me/EflMJ

>The PoneAI drive, an archive for AI pony voice content:
derpy.me/LzRFX

>The /mlp/con live panel shows:
derpy.me/YIFNt

>Clipper’s Master Files, the central location for MLP voice data:
mega.nz/folder/jkwimSTa#_xk0VnR30C8Ljsy4RCGSig
mega.nz/folder/gVYUEZrI#6dQHH3P2cFYWm3UkQveHxQ
mirror: derpy.me/c71GJ

>Cool, where is the discord/forum/whatever unifying place for this project?
You're looking at it.

Last Thread:
>>38605052
>>
FAQs:
If your question isn’t listed here, take a look in the quick start guide and main doc to see if it’s already answered there. Use the tabs on the left for easy navigation.
Quick: derpy.me/FDnSk
Main: derpy.me/lN6li
(Main doc currently down, working on replacement)

>Where can I find the AI text-to-speech tools and how do I use them?
A list of TTS tools: derpy.me/A8Us4
How to get the best out of them: derpy.me/eA8Wo
More detailed explanations are in the main doc: derpy.me/lN6li

>Where can I find content made with the voice AI?
In the PoneAI drive: derpy.me/LzRFX

>I want to know more about the PPP, but I can’t be arsed to read the doc.
See the live PPP panel shows presented on /mlp/con for a more condensed overview.
derpy.me/pVeU0
derpy.me/Jwj8a

>How can I help with the PPP?
Build datasets, train AIs, and use the AI to make more pony content. Take a look at the quick start guide for current active tasks, or start your own in the thread if you have an idea. There’s always more data to collect and more AIs to train.

>Did you know that such and such voiced this other thing that could be used for voice data?
It is best to keep to official audio only unless there is very little of it available. If you know of a good source of audio for characters with few (or just fewer) lines, please post it in the thread. 5.1 is generally required unless you have a source already clean of background noise. Preferably post a sample or link. The easier you make it, the more likely it will be done.

>What about fan-imitations of official voices?
No.

>Will you guys be doing a [insert language here] version of the AI?
Probably not, but you're welcome to. You can however get most of the way there by using phonetic transcriptions of other languages as input for the AI.

>What about [insert OC here]'s voice?
It is often quite difficult to find good quality audio data for OCs. If you happen to know any, post them in the thread and we’ll take a look.

>I have an idea!
Great. Post it in the thread and we'll discuss it.

>Do you have a Code of Conduct?
Of course: 15.ai/code

>Is this project open source? Who is in charge of this?
derpy.me/CQ3Ca
>>
File: anchor.png (33 KB, 1200x1453)
33 KB
33 KB PNG
>>38718598
Anchor
>>
File: pinkie pie but better.jpg (242 KB, 1560x2100)
242 KB
242 KB JPG
when are we getting models for the shadowbolts
>>
File: Spoiler Image (1.15 MB, 1920x1080)
1.15 MB
1.15 MB JPG
New character added:
Verity, Director
>>38713248
Moonlight Raven and Sunshine Smiles

Alright, now I just need ideas for just few more canon ponies to add to 1st and 2nd bench on the right and the crowd will be done.
>>
File: rdash.jpg (1.62 MB, 1400x2234)
1.62 MB
1.62 MB JPG
>>38718601
Rainbow Dash - When Love Takes Over

https://u.smutty.horse/mhqcdembzte.mp3
>>
>>38710346
I got this error yesterday, and today, too. I tried a fresh notebook, no difference. Lowering the batch size doesn't seem to help either. I also tried muddling my way around changing paths. But it worked fine for weeks so I don't have a clue what broke.

For now, I've lowered the checkpoints count, and that lets me slowly train, but it's not an ideal workaround with it adding time to the training and eventually bugging out after a few steps, anyway. It's weird, because the model it produces on the last step does actually work.
>>
>>38718712
Lyra playing a harp with an actual hand and bonbon next to her looking horrified at the monstruosity of a pony with hands.
Octavia grilling for celestia's sake and a shirt that says luv me cello, luv me grill, 'ate glowies, simple as.
Sci-twi nerding out with a couple of notepads filled with gibberish next to her and Sunset Shimmer looking frustrated at the loss of her seat.
Granny smith on the right of the stage with a cart selling "complementary" cider.
Chrysalis sneaking over the wall to hear how to make AI without paying (based).
Rainbow Dash with a trenchcoat and sunglasses trying to keep the fact that she's into the nerd stuff quiet.
An anonfilly plushie next to or on the head of any of the actual anons.
Woah Nelly (the WIDE mare on the princess dress episode).
A fat earth pony with a fedora and shirt ready to ask if celestia AI is god forbid a molester with some of spilled spaghetti at his around.
>>
Looks like Google has stepped up to the plate and made a model comparative/competitive to Dalle2

Google's "Imagen" AI
https://youtu.be/HyOW6fmkgrc
https://imagen.research.google/

Just like Dalle2, this looks extremely promising. One thing in my mind is that because it's not OpenAI's, I'm curious if it'd be easier/possible to access it's source, or it being just as if not more closed.
>>
>>38719088
>IsADirectoryError: [Errno 21] Is a directory: '../wavs/'
How does your training/validation text looking?
Is it's format different than this example:
wavs/00_00_00_Character_Serious__You call that a knife.wav|You call that a knife?;
>>
>>38717899

Wow, something where my knowledge comes in handy. Alicia is a poster on the /mlp/ discord and full of lulz and pmvs. She has a handicap of some kind (cerebral palsy I think?) and hangs out there. I doubt this was aimed at our BGM
>>
>>38718712
kekked at that snowpony ice one! Very creative!
>>
Clipper you around?
>>
>>38719622
He's usually away on weekends.
>>
>>38719622
I'm here, what do you need?
>>
File: Ohne Titel(2).png (1.79 MB, 1920x1080)
1.79 MB
1.79 MB PNG
>>38718598
https://www.youtube.com/watch?v=JeoZgAhfMeM

Made another one
>>
File: 1605975010650.png (242 KB, 800x787)
242 KB
242 KB PNG
>>38720407
>Ohne Titel(2).png
>>
File: SEX WITH TRIXIE.png (66 KB, 600x652)
66 KB
66 KB PNG
>>38720407
Absolute kino.
>>
>https://openvoice-tech.net/index.php/Pony_Preservation_Project
>This page was last edited on 30 November 2021, at 21:15.
Anyone feeling to fill in this dudes page with were to get PPP datasets, models and what not?
>>
15.ai is on the front page of hacker news right now
>>
>>38720509
That entire wiki is defunct, no one has made edits in months. Just use the 15.ai Wikipedia article if you have any (sourceable) information for the PPP.
https://en.wikipedia.org/wiki/15.ai
>>
>>38720523
I’m surprised that article hasn’t been vandalized by the ubercucks. But then again wikipedia probably has a really good anti vandal system.
>>
>>38720521
>this AI works best with highly pitched voices?
So should we train South Park too?
>>
>Page 9
>>
Reposting from the Fimfiction thread:
>>38717430
https://www.youtube.com/watch?v=YQrO6G9HSYQ
https://u.smutty.horse/mhqepskeort.mp4
>>
>>38719344
>google
yeah, I wouldn't hold my breath
>>
>>>/g/87329616
Kek
>>
>>38719344
https://github.com/lucidrains/imagen-pytorch
Whenever a big new paper comes out using transformers, just wait a few days and check lucidrain's github.
>>
>>38721658
is this useable right now or is it pointless/too heavy
>>
>>38718598
Which will reach the finish line first?
Text to image or AI voices?
>>
>>38721670
Depends on where you draw the finish line. Spoiler:
There is no finish line.
>>
>>38721670
>>38721675
To be honest, AI voices have pretty much hit the "finish line" in terms of sounding generally good, although it only really applies to the large datasets (see Twilight) and, y'know, there are definitely still flaws.
>>
>>38721670
>>38721675
>>38721678
The thing about AI is that getting 90% of the way is easiest part. Getting 99% of the way there is exponentially harder, and 99.9% of the way is even harder than that.
>>
>>38721678
Exactly, you could make the case that they're "already there" but of course there's still room for improvement. That principle, I imagine, won't change in the foreseeable future. If all of the current flaws were fixed, stuff like the emotional control being stilted as fuck would become the next perceived problem. Fix that and something else would resurface.
>>
>>38721667
It looks like the only things missing from the repo are the pretrained unets. You can probably ask the LAION people for those since it looks like they trained it.
I don't know how big the models are. It doesn't look huge, so if you have a good gpu, you can probably train it. It uses a 1GB T5 model by default, and the unets are probably not much bigger.
>>
I'm currently writing about Talknet, GPT-PNY, and the various image generation scripts posted to the thread. I'm not that familiar with the technical details of these, but I'd like to go over them since they're relevant to the direction of the thread.
If you have anything you want me to mention, please let me know. I can compile it into a script, but I'll need at least bullet points. Note that I'm focusing specifically on the technical details. I'm open to non-technical details too, but they might need to go into another section.
If you want me to bring up anything technical that's not about Talknet, GPT-PNY, or image generation, I'm open to that too, especially if it's something I didn't work on. I'll see if I can work it in.
>>
>>38721658

Interesting. So if I'm understanding this right, this is an independent remake of the guts of imagegen, but there's no pre-trained stuff yet. The positive being we could potentially train our own censorship-free models? But the downside being that's a hell of a lot of work and until someone trains something there's not much to play with. I don't know about running locally, but the comments there seem to suggest it can train and run on collab at least.
>>
>>38722152
>So if I'm understanding this right, this is an independent remake of the guts of imagegen
That's correct
>but there's no pre-trained stuff yet
There might be. Of the three required models, only 1 is pretrained and available. It was developed within the LAION community, and some people there might have already trained the other two required models.
>The positive being we could potentially train our own censorship-free models? But the downside being that's a hell of a lot of work and until someone trains something there's not much to play with.
Correct.
>I don't know about running locally, but the comments there seem to suggest it can train and run on collab at least.
I'm reading through some of the LAION community messages on training this. If you want to train it to the same extent that Google did, it will take too much compute. But the amount of compute you use should scale down with the dataset you use. Google used approximately 4.5 years of RTX 3080 Ti time for 800M images. If we use 1M images, we would want to update the model to use 800x fewer parameters, and we would need approximately 2 days of RTX 3080 Ti time for training. It should be doable for us.
>>
>>38722362
>If we use 1M images, we would want to update the model to use 800x fewer parameters, and we would need approximately 2 days of RTX 3080 Ti time for training.
Correction: we would want to decide how much data to use based on how much time we're willing to spend training. If we're willing to spend 2 days of RTX 3080 TI compute, we would want to train the model on 1 million images. If we're willing to spend 10 days, we should train it on 5 million images.
We would want to play around with smaller models to see exactly how dataset size should scale with how much compute we have, but it would be something like that. The principle of scaling the dataset with amount of compute available comes from https://arxiv.org/pdf/2203.15556.pdf which was designed for language models. We would want to test it to see if it works for image models too.
>>
>>38720612
These days, it's extremely difficult to vandalize any Wikipedia article and keep it that way for more than a few minutes.
>>
>>38719088
I agree with >>38719444. Look at /content/hifi-gan/training.txt and /content/hifi-gan/validation.txt and make sure that each line is a valid file. There should not be any blank lines or lines with just a folder and no file (like "../wavs/")
>>38722123
I read through the current progress in the doc, but I'm still not sure exactly what you're looking for. Technical details of how/why the models work? What can/will be done with them? What kind of audience are you targeting? (e.g. how much technical knowledge do you expect people to already have?)
If you're interested, I could write some bullet points about why TalkNet's design allows for controllable TTS and ngrok's does not. I could also try to summarize recent advances in text-to-image generation (e.g. why diffusion is good, advancements with CLIP, Latent Diffusion, DALL-E 2, Imagen, etc.), though I am not an expert.
>>38701487
>>38713884
>>38719078
Nice! Quite a lot of RD songs. Wonder which pony has the most songs in total
>>
>>38719444
All my wavs/, .wav and | are in the right places as far I could tell, but you're probably right. There's probably a symbol I can't find that's causing problems. I should probably start using one of the pieces of software specifically for this, instead of just using a text editor.
>>
>>38720521
It was surprising to see how much orange site has changed in just a year. I want to believe it's merely uberduck shills shitting things up, but maybe they're finally reaching their September.
>>
>>38722374
>'m still not sure exactly what you're looking for
Anything that indicates the significance of TalkNet / GPT-PNY / the image generation scripts, whether that's significance for the PPP or significance for AI more broadly. It could be improvements over what was available previously, anything unique to the PPP's version of these things, problems that had to be overcome, what directions they open up, or anything strange/surprising about them. For TalkNet and GPT-PNY, if you have any details on what it took to train them, that would be good too.
>If you're interested, I could write some bullet points about why TalkNet's design allows for controllable TTS and ngrok's does not. I could also try to summarize recent advances in text-to-image generation (e.g. why diffusion is good, advancements with CLIP, Latent Diffusion, DALL-E 2, Imagen, etc.), though I am not an expert.
That would be great. I'm planning on reading through some of these papers too so I can better understand their significance. If you can summarize what you know, then hopefully between the two of us we can get all of the important parts.
>What kind of audience are you targeting? (e.g. how much technical knowledge do you expect people to already have?)
The target audience would be very broad, so I'll be assuming low technical knowledge for most of it, but I'll be including more dense technical information where it's useful and where I think less technical people will tolerate it.
I'd prefer it if your bullet points were as technical and as accurate as possible. I'm not going to make the presentation overly technical, but I want to make sure that I understand what's going on very precisely if I talk about something. If we mention it in the presentation, there's a decent chance that it'll come up in the Q&A, and I want to make sure we have something intelligent to say for every question, even for ones where we need to defer.
>>
>>38722424
>>38719444
You were right! There were unnecessary blank lines at the end of both lists causing the breakage. And here I was, blaming the packages instead.
>>
>>38720407
Very nice. I like the spacey vibe from the voice effects.

>Equestria Portal opens November 17th
>Massage a mare to arouse her, also play music to relax.
>Twilight Sparkle is likely into humans
Where are those prompts generated from?
They look fun and pony aware.
>>
>>38722434
You're not wrong. Literally in that thread:
>peterlk 13 hours ago | prev | next [–]
>I have no involvement in either of these companies, but I'll mention that this seems like a beta version of uberduck. Personally, I think uberdurk is awesome, and probably worth a look
These faggots are so pathetic.
>>
File: ubercucks.png (6 KB, 1150x65)
6 KB
6 KB PNG
>>38722994
4got pic
>>
Live in ~1 hour. Panel planning and scriptwriting.
cytu *dot* be/r/PonyPreservationProject

I'm having difficulty getting everything finished for the mega compilation, I'm not sure if I'll be able to finish editing and rendering ~45 hours of content in time. Today (for me) will be mostly dedicated to determining if a solution can be found. If not, I'll likely have to cancel it.
>>
>>38722461
>>38722424
I quoted the wrong post.
>>
MARE

https://u.smutty.horse/mhqyjlrctfc.mp4
>>
>>38723686
>fluttershy's portion
>>
>>38723686
millions and millions
>>
I can't wait for somebody to make a talkbot neuronal network fed with /tg/ and /fimfic/ so I can finally have some creative discourse about the MLP episodes. It would be even better if the bot itself could see and hear the episode itself instead of basing all its information on what others tell it. Would regardless be interesting to see how 2 databases combine together and having to be selective with how you teach it. Too much /mlp/ would ruin any good episode discussion.
I would KILL for a 24/7 bot that can come up with new takes on on episodes, books, movies, but it obviously depends where its getting its education from.
>>
>>38723895
You'll have the exact same discourse that exists currently in PPP: One side of the fandom pushing to include as much data as possible to increase the training on the model, and one side pushing to include only specific content to reinforce their head canons.
>>
>>38723916
The only things that will get done are the things people choose to work on. It doesn't matter who argues the loudest.
>>
>>38718601
>>38723180
General summary of today's PPP panel discussion:
Script for the technical segment is mostly finished. The main thing left to add is a short vision of the future/contributing segment, which will be followed by practise and polish.

Got 10 out of ~45 videos rendered for the mega compilation. Still not sure if we'll be able to do all of them but we're going at a decent rate.

Planning docs:
https://docs.google.com/document/d/1_kOLXFhHNF0vuiKbJXLd82YbpvLn0LDel4uG429cyis
https://docs.google.com/document/d/1NJz3rpw7Amxd_B38_o5oJdV0TIUezXlrNx8Hwy7k5eE

Slides:
https://docs.google.com/presentation/d/1udxmC5cqQejEfygn2Uxs7D4EPwwQD74JnrhYSjP1pS0
https://docs.google.com/presentation/d/1uXbQUTnm8YvNhziLPOdiyDuceFg_hAFvrrSmHAGHySE

Today's stream:
https://youtu.be/kViYwXjMC80

Full playlist of all streams:
https://www.youtube.com/playlist?list=PLX9nDSq9VgBMeIi8jzTrJGFttLZH_1rOO
>>
>>38724242
Correction: the main thing left to add in the technical segment is that future/contributing segment plus >>38722461. Then slides, then practice & polish.
>>
>>38724242
Can you upload an example blender project packed or unpacked? If it's just rendering then I might be able to make it faster. If blender chokes on turning an image into a video then I could use another program to make a video with a frame rate from an image. Then just the audio could be exported from blender and added to the video. Although I understand if you don't want to work on it anymore.
>>
>>38723916
>>38723895
Honestly, if I could pack the ai model with BEST quality quest writers stories and combine it with the top tier quality drawfags work and just have unlimited MS Paint adventure with whatever the fuck I want the ai to DM for me, I will be set golden for life.
Like you could have future ai cure cancer, create a functional space elevator and even make diplomatic contact with the alien civilizations I would still say "nah fuck that, smutty stories with sexy mares is the historic highlight of entire ai development".
>>
>>38723895
This isn't how neural nets work. There is no NN AI currently in existence that can ingest hundreds of hours of high def video and audio and then spit out text expressing an opinion on it. For that matter, there's no NN that can actually create something fully original full stop. GPT works by using an insane amount of statistical modelling to predict which character in a sentence comes next, give what's come before. That's why you always have to start with a "prompt." "AI", as it exists now, cannot actually generate novel content, despite giving the appearance that it can. All output is simply from source material it has access to, rearranged and paraphrased. If you fed an AI a ton of scientific knowledge and then started prompting it it could regurgitate what you had provided, but it would never present you with some grand new discovery.

The content just LOOKS novel, because it's pulled from so many different sources and recombined cleverly. Fundamentally, AI is just the world's best search engine. It's as if you hired a dozen interns to research something, report back to a head intern, and then had him present the findings in a nice two page summary. Very convenient, but they're just researching, not creating anything new. They can tell you what 1000 other people have said about "Bridle Gossip," but they can't say anything new.
>>
>>38724589
>The content just LOOKS novel, because it's pulled from so many different sources and recombined cleverly.
You could argue that this is also the case for humans since no idea is fully original. Although transformer models currently cannot filter the nonsense in their outputs as long as they make enough sense in the limited context they work with, which is why they can sound rather clever when you engineer your prompts well enough, but also be fully convinced that horses have five legs if you say the correct words.
>>
>>38724647
>You could argue that this is also the case for humans since no idea is fully original.
NTA but while you could argue that, I find that outlook to be overly reductive, not to mention provably false.
At some point in time none of the ideas we currently have existed. For them to exist now, someone must've actually created them. I'd argue that the process is a fair bit complicated and that the creation of new ideas is more of an iterative method than we think but that it still doesn't mean the end result is never original. It only means it's very hard (if not outright impossible) to find the precise moment where an individual's thought process created an idea that didn't exist prior. Even many of the ideas we consider cliche' would be incomprehensible for, say, ancient Greeks or Sumerians.
>>
>>38724718
NTA, but I think the point was that the same arguments you're making in favor of human creativity can be applied to neural network creativity. To say that neural networks have no creativity is overly reductive, and the claim that neural networks only rearrange and paraphrase training data is provably false.
Then again, I don't know why people place so much sentimental value on "true" creativity. AI text generators are often, though not always, more creative than me in ways I care about. When I'm using it, that matters infinitely more to me than whether its creativity is "true" in some useless sense.
>>
From what I've been seeing floating around in demos lately, is it safe to say that commercial voice transfer models have absolutely blown things like Controllable TalkNet out of the water in terms of quality? I wonder how long it'll be before more public domain models of that caliber become available
>>
I was right about Lamda algebra and ram I wonder what else I've been right about
>>
>>38725303
Commercial models likely benefit from more data/training time, a dedicated team, and so on. Also, TalkNet is not so much a voice transfer model as it is a TTS model that can be used to transfer voices. So, models that go directly from voice to voice (end-to-end models) have advantages. For example, they can completely avoid phonemes, which allows for training on untranscribed datasets (there is much more untranscribed data than transcribed data, and being able to leverage more data leads to quality improvements).
Was there any demo in particular that impressed you?
>>
>>38722461
Thanks, that was helpful. Just to be safe, what's the deadline for these bullet points? I don't want the hold the preparation process back.
Regarding diffusion, I realized that everything I can talk about is probably too technical. For example:
>How does diffusion compare with other generative modeling approaches?
>- AR models: slow to sample from, enforcing causality restricts model architectures
>- GANs: mode dropping, convergence difficulties
>- Normalizing flows: requiring invertibility results in large models
>- Diffusion: maximum likelihood training is stable and doesn't drop modes. Few restrictions on model architecture (main one is that model input size = output size = data size). Produces an entire sample at once, can shorten noise schedule to sample from in fewer steps.
If there was a pony diffusion model, there would at least be samples to show, but I don't know of any. There might be enough time before the con to train a very tiny model (finetune a CIFAR-10 model on 32x32 pony faces), but it doesn't seem worth the effort.
So, I figure it's best to just stick with TalkNet. (I don't know enough about GPT-PNY to comment.)
>>
>>38725738
we're so constrained by hardware it's ridiculous. If we didn't get alms from google even in this reduced state we wouldn't be getting anything done.
>>
>>38725763
>Just to be safe, what's the deadline for these bullet points?
Thursday, end of day PST.
I'll need to have the script and slides mostly finished by Friday end of day and fully finished by Sunday.

>I realized that everything I can talk about is probably too technical.
Don't worry about it being too technical. I can take care of translating it to something more accessible. For example:
>GANs
Training is unstable, and it's not good at dealing with many different styles. TPDNE is a good example showing mode collapse, where all of the images look like they're of the same style despite being trained on a wide variety of faces from derpibooru.
>Normaling flows
Models need to be bigger to get the same level of quality.
>Diffusion
Training is stable, it can create images with many styles, it's easy for AI anons to make changes to the model, and there's a slider for trading off speed for quality even after the model is trained.
>AR models
It takes longer to train and run than other models, and it's harder to fix consistency problems in models.
This is all good information, by the way. Thank you.

I don't think we have any pony-specific image generation models, and it would be too much of a time crunch to try to create one before the panel. That's okay though. We can talk about non-pony models since those still produce interesting results when trying to generate ponies and since anons here have started playing around with them. Some chunks of the panel are about the general direction of the PPP, and that includes things we haven't yet done. It's some mix of educational info, hype, and PPP history. We'll definitely want to discuss TalkNet, and I'd like to discuss the image models too. I'll try to get comments on GPT-PNY from Delta.
>>
>>38723180
Live in ~1 hour. Panel planning and scriptwriting.
cytu *dot* be/r/PonyPreservationProject

>>38724298
Gothic and I have already been trying to share the packed projects between ourselves and have both got similar slow results. We also found that there'll often be missing assets in the projects despite them all being ostensibly packed. We've been able to resolve these issues between ourselves but I suspect you'll likely run into the same issue. If you can find any settings to make it go faster, that would be very helpful.

Here's where I'm storing packed blend files. Gothic already has all these and is working on rendering them. If you can find any changes to make that finish rendering in less than 2-3 hours then that'll be an improvement on where we currently are.

If you'd like, we can also try to work with you over the stream to get everything working if you want to help do some of the final renderings.
>>
>>38726590
>forgot link
https://mega.nz/folder/pBFQ2Jga#OdcSqhLNymkdkhOvCBHcRg
>>
MARE N*****

https://u.smutty.horse/mhrkdkiuevs.mp4
>>
File: Spoiler Image (1.21 MB, 1920x1080)
1.21 MB
1.21 MB JPG
>>38726590
General summary of today's PPP panel discussion:
Synthbot was helping me touch up the information on softprompts while he also worked on part of his script.
Also the crowd in panel art is pretty close to getting finished.

Planning docs:
https://docs.google.com/document/d/1_kOLXFhHNF0vuiKbJXLd82YbpvLn0LDel4uG429cyis
https://docs.google.com/document/d/1NJz3rpw7Amxd_B38_o5oJdV0TIUezXlrNx8Hwy7k5eE

Slides:
https://docs.google.com/presentation/d/1udxmC5cqQejEfygn2Uxs7D4EPwwQD74JnrhYSjP1pS0
https://docs.google.com/presentation/d/1uXbQUTnm8YvNhziLPOdiyDuceFg_hAFvrrSmHAGHySE

Today's stream:
https://youtu.be/zK9cwJ2Jgdk

Full playlist of all streams:
https://www.youtube.com/playlist?list=PLX9nDSq9VgBMeIi8jzTrJGFttLZH_1rOO
>>
So I'm having an issue with my audio in Adobe Animate flatlining. The WAV plays fine in Discord and Audacity, but in Adobe Animate, it's totally silent. Event is set to stream and I tried Googling with no help. Anyone possibly have more information?

https://u.smutty.horse/mhrmhckdkuk.wav

Are there any properties of WAV files for why this one doesn't work but others I already imported do?
>>
>>38728280
There could be plenty of reasons. Try just re-exporting it from Audacity and using the exported clip in Animate instead.
>>
>>38726590
You still need music for that panning across the /mlp/ map thing?
>>
>>38728280
That WAV file's encoding is 32-bit floats, which is not very common. Try opening it in Audacity, exporting as WAV, and selecting "Signed 16-bit PCM" for "Encoding".
>>
Was inspired to create a Dalle-mini mare art by another thread.
Seeing as it is AI related and thus relevant to PPP, I'll share her here too.

Meet "Juicy Peaches".
She currently has a stuffy nose.
Allergies? Cold? It's unclear.
Anyways, enjoy the cute mare.
>>
>>38728552
cute mare
>>
>>38728280
Adobe Animate requires you pawn off your first born if you pay for it legally, and it can’t just handle most audio formats?
>>
>>38728552
nothing like a cute mare to end the night
>>
>>38728552
Are you going to put that on Derpibooru?
>>
I just realized that TalkNet was deprecated in April and completely removed from NeMo two weeks ago, along with models such as FastSpeech and Glow-TTS (https://github.com/NVIDIA/NeMo/pull/4082). From https://github.com/NVIDIA/NeMo/discussions/4178:
>We recently unified the TTS models' data loading under TTSDataset, which breaks support for older checkpoints that depend on model-specific Dataset classes. As a result, we've decided to deprecate some older and lower-quality models to focus on and update a smaller set of flagship models.
Now the only models that support pitch inputs are FastPitch and Mixer-TTS, both of which use phoneme or character average pitch. I think this will make it harder for the model to be expressive.
For example, when Pinkie sings "the sun was going do-o-o-own", the {AW} sound lasts for 1.5s. Since TalkNet doesn't average pitch, the model receives 130 pitch values. FastPitch and Mixer-TTS only receive 1 or 2 pitches, though, which means that the model has to fill in the pitch contour. If it isn't powerful enough to do so, then the output will likely sound monotone or off-key.
Of course, this isn't an inherent limitation and can be removed by modifying the code. Doing that would prevent you from warm-starting with the pre-trained checkpoints, though.
>>
>>38728295
We've got some fairly generic "adventure" type music at the moment, but I'm sure it could be improved if you want to give it a go. Here's the panning video, it'll crossfade to Gothic's panel crowd drawing once it's finished, which will then be the backdrop for ~five minutes of intro clips.
https://u.smutty.horse/mhraetgataj.mp4
>>
File: 1655281859179.png (204 KB, 604x760)
204 KB
204 KB PNG
lol
>>
File: google-research t5x.png (59 KB, 520x450)
59 KB
59 KB PNG
https://github.com/google-research/t5x
Bros, how can we use this to make more ai ponies?
>>
>>38728581
>>38729218
Indeed, glad yous like her.
>>38729245
I don't see why not. I'll probably export a higher definition vector version in a bit.
>>38729557
If any of us can run it that is. 1 Trillion parameters sounds very chonky.
>>
File: Spoiler Image (1.22 MB, 1920x1080)
1.22 MB
1.22 MB JPG
Bumping with updated panel art, and at the moment Im 99% fine with the current lineup of the crowd.
>>
>>38726590
Live in ~1 hour. Panel planning and scriptwriting.
cytu *dot* be/r/PonyPreservationProject
>>
File: pewpew.png (177 KB, 800x600)
177 KB
177 KB PNG
So we were talking about some theme song for the panel, how about a parody of There's a Great Big Beautiful Tomorrow? It's very apt, if nothing else:
https://www.youtube.com/watch?v=20FcZDAIAO8
It's the Pony Preservation Project
Keeping ponies safe another day
It's the Pony Preservation Project
Hijacked research to make the good times stay
Man has a dream and that's the start
He follows his dream with mind and heart
And when it becomes a reality
It's a dream come true for you and me
So join the Pony Preservation Project
The ride doesn't have to be so gray
Join the Pony Preservation Project
Make more mares today!
>>
>>38730431
I feel like it would be really neat if people started to make original songs but in the style of the 30s~60s music, creating a neo-retro style of musical subculture.
>>
File: ezgif-5-7f0f08211a.gif (27 KB, 125x125)
27 KB
27 KB GIF
Was bored before going to bed and had this TalkNet Twilight song stuck in my head, so I found a suitable gif and synced it best I could via ezgif. I quite like the result.

Twilight is blinded by the lights:
https://u.smutty.horse/mhrvofwnosb.mp4
>>
>>38730790
Perfect
>>
We'll be right back!
https://u.smutty.horse/mhrwpuewquf.mp3
>>
>>38730096
General summary of today's PPP panel discussion:
More scripts, more slides.
Rendering of the mega compilation videos is about halfway done, looks like it'll just about be all ready at the current rate.

Planning docs:
https://docs.google.com/document/d/1_kOLXFhHNF0vuiKbJXLd82YbpvLn0LDel4uG429cyis
https://docs.google.com/document/d/1NJz3rpw7Amxd_B38_o5oJdV0TIUezXlrNx8Hwy7k5eE

Slides:
https://docs.google.com/presentation/d/1udxmC5cqQejEfygn2Uxs7D4EPwwQD74JnrhYSjP1pS0
https://docs.google.com/presentation/d/1uXbQUTnm8YvNhziLPOdiyDuceFg_hAFvrrSmHAGHySE

Today's stream:
https://youtu.be/2eOaFlDNwF8

Full playlist of all streams:
https://www.youtube.com/playlist?list=PLX9nDSq9VgBMeIi8jzTrJGFttLZH_1rOO
>>
>>
>>38731489
why does she have paper in her tail
>>
>>38731493
thats part of the treatment
>>
>>38731489
Despite being a quadruped, that leg holder thing does not look at all comfortable for her.
>>
>>38731489
Because she dyes her mane and tail. She uses the paper to keep the dye properly separated. She also lies to fillies.
>>
>>38731690
>She also lies to fillies.
I don't want to invoke that ancient argument but she obviously doesn't. Spoken dialogue > silly animation joke. She also had the same manestyle as a filly.
>>
>>38731706
They put it there for a reason. And what's stopping her from dyeing her mane as a filly?
>>
>>38729249
There are some tricks in NLP for augmenting the input layer of a trained model. We can't use the exact same tricks for speech, but they at least show that it's feasible to update the input layer much more cheaply than training a whole new network.
That does suck though. It means the TalkNet line of models will be much more difficult to update than the TTS line of models.
>>
>>38731725
I was halfway into writing a long-ass post explaining why you're wrong but fuck it, I won't take the bait more than I already did just this once.
The Anons of the past didn't spend a thread on this exact issue for nothing.
>>
>>38731706
Maybe her colors faded, and she's dying it a brighter shade of her natural colors. Maybe the colored strands didn't grow in whatever proportions she hoped for. Just because she didn't have a reason to dye her mane when she was a filly doesn't mean she doesn't have a reason to dye it now.
>>38731750
Link the thread. I have yet to see a convincing reason to totally disregard what's obviously on screen.
>>
File: Twilight_mane_argument.png (953 KB, 4136x3952)
953 KB
953 KB PNG
>>38731766
It's a relatively famous cap, I'm even in it. I didn't actually read through that it since 2014 but I imagine that every single possible argument ended up being brought up so if you want to go delve in the catalog, >>16446708 is where it starts.
There was no consensus in the end though, which is a part of the reason why it became a board meme.

As for me, I still maintain that you shouldn't take as canon anything that was clearly, without a doubt, put into the show as joke (no, Rick and Morty aren't "canon" either) but apparently that point of view isn't even close to a universal opinion.
>>
>>38731766
>Maybe her colors faded
Perhaps multicolored manes need different kinds of conditioner for each color to be at it's healthiest/brightest.
Twilight strikes me as the kind to do that just so each of her hair colors' health and level of pigment intensity is consistent with each other.
A matter of personal organization of the body rather than a matter of beauty.
>>
>>
>newfags genuinely got baited into a genuine Twilight's mane argument in the year of 2022
Never change, /mlp/, never change. I bet some Anons will be having this argument in 2222, too.
>>
>>38731742
>There are some tricks in NLP for augmenting the input layer of a trained model.
I'm not familiar with these tricks, but you do bring up a good point. Instead of starting from scratch, you could do model surgery to feed non-averaged pitch embeddings into the pre-trained model as an additional source of information. Finetune for a while on LJSpeech (or whatever the dataset is), and you should end up with a base model that can be used for further training.
>>
>>38729469
Link to tweet for posterity: https://twitter.com/SwiftOnSecurity/status/1536908941052891137
>>
>>38726590
Took a look at the blend files and there's not any glaring issues with your settings. When running it I noticed that the blender animation sequencer only uses 1 thread. It's weird because the 3d renderer can use multiple theads/gpu. There's workarounds though like
https://github.com/sciactive/pulverize
which calls the blender api in multiple processes using multiple threads. More of a hack than a solution although if the alternative is using a single thread for 3 hours then it might be worthwhile to the linked blender script or another video editor entirely. I'm a bit late to the party since my car broke down so sorry. Although it seems like you and Gothic are getting through it somehow. It's just something to keep in mind if there's going to be a next time for something big like this.
>>
>>38731182
Why is this site *still* under construction?

https://voicechasers.com
>>
>>38732159
I'd mentioned that during a stream, though was under the impression that multiple threads were CPU specific, and so the speed wouldn't differ much of using 1 GPU thread vs multiple CPU threads. Am I mistaken and this plugin actually does allows all GPU threads for VSE?

If so, that'd prove incredibly beneficial, and almost crucial for how much me and especially Clipper use it.
If not, then I'm not convinced it'd be worth adding.
>>
File: 1652241572541.png (2.23 MB, 1768x3300)
2.23 MB
2.23 MB PNG
>>
>>38730096
Live in ~1 hour. Panel planning and scriptwriting.
cytu *dot* be/r/PonyPreservationProject
>>
>>38720523
Hey can someone with a Wiki account add the article to the MLP WikiProject? I would do it but I don't know how to
>>
>>38726590
>>38732159
>>38732998
https://www.youtube.com/watch?v=LJQptYvXAnw
https://pypi.org/project/bpsrender/
I've spend few hours looking into the multi thread render and got this code to working (to some degree).
First need to install bpsrender with the line "pip install bpsrender", than make sure you have installed the 'ffmpeg', after that, while not explained in their installation guide you MUST link the directory folder of 'ffmpeg' and the directory of the blender.exe you are using e.g. "C:\Program Files\Blender" to the System PATH variable environment (more instruction on how to do it https://superuser.com/questions/1099259/adding-ffmpeg-as-environment-variable).

Once all of the above is done, what you need to do is
1) go to your ".blend" file, make sure EVERYTHING is set as it should be for a normal rendering (GPU GB usage is set, missing external files are re-linked, the path is chosen, the output format codec and other is the one you desire e.t.c.), you than save and close that blender file.
2) press right shift and click empty space next to the '.bend' file, and choose "open command window here" (or alternative open the cmd from start and than use teh directory options to go to the location of the '.blend' file)
3) type "python -m bpsrender -h", it will show you all the options that come with the bpsrender rendering outputs and descriptions for them
4) type "python -m bpsrender YourBlenderFile.blend --dry-run" to first test if the bpsrender works at all
4) type "python -m bpsrender YourBlenderFile.blend -w 2" to make it use "2" cpu threads.
PS I would recommend trying "-w 2" and than upping them by one to see what you can get away with as it will create the -w "number" of blender program instances with that file project opened up, I've made an mistake of picking 4, and since the project ive test it on is 11GB in size when loaded it bloody crashed my pc with no survivors.
>>
>>38734173
Awesome, I'm definitely gonna give this a try when I next do a project that requires Blender's VSE.

>press right shift and click empty space next to the '.bend' file, and choose "open command window here" (or alternative open the cmd from start and than use teh directory options to go to the location of the '.blend' file)
Another trick I've found which is super convenient for opening a cmd file in the right directory, at least for Windows 10, is just going to the address bar of the file explorer whilst in your chosen location, replace what's in there with "cmd" and press enter, it'll open up a command prompt set to that location by default.
>>
>>38732998
If the encoding is happening on the gpu then it'll be fast and it won't matter how many cpu threads you use. Although if it's doing cpu rendering on a multithreaded cpu then adding say 4 to 10 threads would make a huge difference in encoding time. Blender's video editor doesn't use the gpu and doesn't use multithreaded rendering. The tool that I linked does multithreaded cpu rendering by for each thread loading a whole copy of blender, rendering a range of frames, outputting each range of frames to a separate video file, and concatenating the files without reencoding. This blows up memory quickly as the number of threads increases. The thing that bothers me the most about this is that blender supports video encoding through ffmpeg and ffmpeg can do either gpu encoding or multithreaded cpu encoding. So adding the logic for fast and memory efficient video encoding would be relatively easy. Blender's ffmpeg work is centered in source/blender/blenkernel/intern/writeffmpeg.c so I could look into adding good video export although it would involve a recompile to apply it. It's possible that a plugin could be developed instead which wouldn't involve a recompile although it's been a while since I've written a blender plugin.

>>38734173
That tool looks more maintained then the one I linked which uses php kek although it seems like it uses a similar methodology and like you say the memory blows up quickly when adding more threads. I'll take a 2-4x encoding speed increase although it'd be nice to take advantage of the gpu or for example 16 threads.

>>38734217
That's a nice trick. I didn't know about that.
>>
>>38734173
*one thing I forgot but it is Horribly (!) important thing I've forgot, when naming your blender files, DO NOT have spaces in the name, as spaces are used to differentiate between the commands, so have either no spaces or just '_' in those spots.
>>
>>38718601
>>38731331
General summary of today's PPP panel discussion:
Most of the script is finished, mostly refinements and slides now.
~2/3 done on rendering and uploading mega compilation videos.

Planning docs:
https://docs.google.com/document/d/1_kOLXFhHNF0vuiKbJXLd82YbpvLn0LDel4uG429cyis
https://docs.google.com/document/d/1NJz3rpw7Amxd_B38_o5oJdV0TIUezXlrNx8Hwy7k5eE

Slides:
https://docs.google.com/presentation/d/1udxmC5cqQejEfygn2Uxs7D4EPwwQD74JnrhYSjP1pS0
https://docs.google.com/presentation/d/1uXbQUTnm8YvNhziLPOdiyDuceFg_hAFvrrSmHAGHySE

Today's stream:
https://youtu.be/ZJQ7-2NYmcs

Full playlist of all streams:
https://www.youtube.com/playlist?list=PLX9nDSq9VgBMeIi8jzTrJGFttLZH_1rOO
>>
>>38729510
>>
The best text generation to teach your neuronal network AI from is /tg/ pre 2012 and then slowly teach it in an educated way about how to review episodes and fanfiction without devolving into shitflinging.
>>
>>38734999
You know, if anyone could have re-discover the pre-"I-fit-my-entire-head-in-my-butt era of the yt mlp analytic channels, you could ript the transcriptions from those videos and turn them into a decent dataset for the text generators to learn from.
>>
>>38734999
Trips checked, but what you're proposing is laughably ridiculous. That said, go ahead and get started, we eagerly await how you plan to develop something like that.
>>
>>38734822
Another thing I've notice, the script will create a "Bpsrender" folder in the same location that stores the "temporarily" work files, sadly you need to remove those after you finish rendering as the script doesn't have an auto-remover.
Also it will ignore your in blender file location destination and just dump it were the project file is, under the same name as the project file.
>>
Some tips for using the GPT-J model: https://ponepaste.org/7480
>>
1 year anniversary of TalkNet: https://desuarchive.org/mlp/thread/37122501/#37134971
>>
>>38736041
Happy Birthday, TalkNet. You are a wonderful model, and you deserve to be happy.
>>
>>38726055
A few hours late, but I posted what I wrote to the doc under the TalkNet and Image Generation sections. Let me know if you need anything else.
>>
Thank you, Talknet, and everyone who contributed to helping me collect hundreds of megabytes of voice sample data only to use the highly trained models to make sexual grunt sounds

Here's to another year of urhing and ouhing, sometimes with punctuation in strange places.
>>
>>38736041
That's amazing how they managed to make something pretty decent while rushing to throw something together to copy /mlp/
Happy 1 year anniversary talknet!
>>
>>38736756
What, did you confuse TalkNet with Uberduck? TalkNet is SortAnon's alternative to Cookie's multispeaker model, while also offering voice style transfer as a feature that 15 doesn't have.
>>
>>38736959
Nah I was just joshing, I guess the joke backfired lol
>>
>>38733838
Live in ~1 hour. Panel planning and scriptwriting.
cytu *dot* be/r/PonyPreservationProject
>>
File: TrixYay.jpg (67 KB, 1073x817)
67 KB
67 KB JPG
>>38736041
yay!
>>
File: screenshot1.png (1.49 MB, 1280x757)
1.49 MB
1.49 MB PNG
>>38737113
Hey Clipper, I'm editing Wikipedia to include a screenshot of content made by the PPP. Would it be okay with you if I uploaded a screenshot from your Tax Breaks video? I think I just need an OK from you so that it falls under Wikipedia's fair use guidelines (or if you included a CC BY-SA 4.0 license in your video description, that should be fine too)
>>
>>38737617
Sure, I'm fine with you using screenshots for the Wikipedia page.
>>
File: s1 fence lean.png (50 KB, 237x186)
50 KB
50 KB PNG
>>38737617
>>
>>38737630
Thanks.

Also, Shadok, if you're lurking this thread, would it be okay with you if I uploaded a screenshot from your "Spy/Demoman" video? (Or if someone has a Twitter account, can you reach out to him and ask him that? Again, just an OK from him should be fine but if he could put a CC BY-SA 4.0 license in the video, that works too.)
>>
>>38737630
Actually, it's better if you do include the CC BY-SA 4.0 license in the video description instead, because a 4chan post is probably not going to fly as evidence with Wikipedia. Just a statement saying something like "I am releasing this video under the CC BY-SA 4.0 license" should be more than enough.
>>
>>38731901
The LJSpeech data is terrible for natural speech. We would train it on pony data, which we have much more of and which a few groups have already found drastically improves TTS quality.
>>
>>38736355
Thank you, this is awesome. I'm running a bit late too, so it'll probably take me until Sunday to turn this into a script. I'll let you know if I have any questions or if I think anything should be changed.
>>
>>38737708
Done.
>>
>>38737737
Thanks, Clipper. You can see the screenshot here: https://en.wikipedia.org/wiki/15.ai#Fandom_content_creation
Please feel free to contribute if you see any incorrect information here.
>>
>>38737617
The visual content of the screenshot is of copyrighted characters, though. Wikipedia will have a hissy fit. And that screenshot doesn't show information like view count or anything. (And what's worse is you have autoplay on!)
>>
>>38737833
I could be wrong, but just because an image contains copyrighted characters, doesn't mean that it's not allowed on Wikipedia. Pretty much every article for any TV or video game character has an image associated with it.
>>
>>38737679
I've contacted Shadok and he has put a statement in the video description. You should be good to go.
>>
>>38737859
Done. Thanks.
>>
>>38737817
Looks good to me.
>>
>>38734916
I thought we were closer to done on rendering the mega compilation. Am I way off?
>>
If I want to engage in this hobby without Google watching me, would it make sense to purchase a server with stupidly high amount of RAM? https://www.45drives.com/products/network-attached-storage/
>>
>>38738372
depending on what you are tying to do? just audio clips you will be fine with 15ai or offline talknet (and also at least 6+ GB nvidia gpu), if you want to train your own models, with the colab script technically there is nothing stopping you for rewriting it so it works off the Jupyter Notebook code.
As for doing any training of text model from scratch, I think renting gpus at the moment the only way to do it (unless you happen to rob entire inventory of a crypto-coinbro).
>>
>>38738454
If only several major forces didn’t coincide to make buying enough GPUs to train this stuff so expensive.
>>
>>38718601
>>38737113
General summary of today's PPP panel discussion:
History/update is basically finished, just a few minor kinks to iron out in the slides. More slides also made for the technical segment.
~3/4 done on rendering and uploading mega compilation videos.

Planning docs:
https://docs.google.com/document/d/1_kOLXFhHNF0vuiKbJXLd82YbpvLn0LDel4uG429cyis
https://docs.google.com/document/d/1NJz3rpw7Amxd_B38_o5oJdV0TIUezXlrNx8Hwy7k5eE

Slides:
https://docs.google.com/presentation/d/1udxmC5cqQejEfygn2Uxs7D4EPwwQD74JnrhYSjP1pS0
https://docs.google.com/presentation/d/1uXbQUTnm8YvNhziLPOdiyDuceFg_hAFvrrSmHAGHySE

Today's stream:
https://youtu.be/AIIep7qPq0o

Full playlist of all streams:
https://www.youtube.com/playlist?list=PLX9nDSq9VgBMeIi8jzTrJGFttLZH_1rOO

>>38738223
We're fairly close to done; should be able to finish it all in time but it's still a bit of a rush.
>>
Where do we go for deliveries and models of others than ponies? I need the voice of the former president Trump for the following:
>You would think that a Twitter account dedicated to pointing out all of the evidences that Donald Trump is the Antichrist would eventually run out of evidences to post.
>Nope! Not even close!
>The connections are quite literally endless! Enjoy, my friends!
Of course any pony, such as Twi, would fit.
>>
The Weeknd - Blinding Lights, featuring Talknet Rainbow Dash
https://u.smutty.horse/mhsuuzuwvob.mp3

I had this stored up on my drive for some time. A bit of the vocal parts, as expected, aren't perfect. But I think it's a significant improvement from my first attempt of using Talknet since it first came out.
Awesome singing Dash art by RyzaDA from the Dash thread!
Hope y'all enjoy Rainbow Dash getting blinded by the lights! In addition, Happy Birthday Talknet!

>>38730790
Was really cool to watch Twilight dancing to the beat! Thank you and appreciate you doing that, Hazy!
>>
>>38738925
This is awesome! Nice work anon.

I was actually intending to do a Twilight one just so I can have a full version to sync Twilight's dances to it. Care to send the voice samples for a Twilight version? If not, I'll just separate as usual. Hopefully I'll get it done soon.
>>
File: filter curve.png (223 KB, 1920x1080)
223 KB
223 KB PNG
>>38738925
Nice. One thing you could maybe try is this Filter Curve effect (pic related) on Audacity with the vocals since Dash's voice is a bit too raspy normally:

https://u.smutty.horse/mhqcbqlansh.mp3 (Off)
https://u.smutty.horse/mhqcdembzte.mp3 (On)
>>
>>38738925
That sounds really good.
>>
File: 2028998.gif (277 KB, 450x377)
277 KB
277 KB GIF
>>38738925
Dashie is now dancing to it. Took me a while to work out how to correctly sync it for such a long sample. This has taught me an effective method learned from experimentation. So thanks ^-^
Note: Image not entirely in sync. Dashie in the video is.

Dashie Dancing to >>38738925's Blinding Lights (Dashie Version)
https://u.smutty.horse/mhswqynndhk.mp4
>>
>>38739460
Also, strangely, it's more in sync downloaded than it is in browser (or like, the sound start is more offset/delayed?).
But maybe that's just Firefox's playback messing with it. Chromefags let me know.
>>
>>38737113
Live now. Panel planning and scriptwriting.
cytu *dot* be/r/PonyPreservationProject
>>
Bump.
>>
Why is this AI Flutters so sad?
>>
>>38739719
Made more progress on the mega compilation rendering and made the first five-minute break video.
>>
>>38738925
This is really good!
>>
File: 1625464409514.gif (3.12 MB, 340x531)
3.12 MB
3.12 MB GIF
>>38738925
>>
File: 1628358456132.jpg (65 KB, 309x266)
65 KB
65 KB JPG
>>38739460
>This has taught me an effective method learned from experimentation
>>
I remember there was a period where Colab basically made it impossible to train a new TalkNet model without paying because of GPU allocation. Is it any better these days?
>>
>>38740587
Yes, it's giving out proper TPUs again.
>>
>>38740669
How are the GPU times before they boot you out for 24 hours? Are you able to get through a solid round of training?
>>
>>38740675
I don't train frequently but as far as I can tell you can use a GPU about as long as you always could, ~6-8 hours. Captcha checks are more frequent so you may have to babysit to make sure it doesn't kick you off prematurely for "inactivity". I think as long as you use checkpoints you should be fine.
>>
>>38740730
Good to hear, hopefully I get time to put together another dataset and try training again.
Speaking of, is there any potential of 15.ai sharing contacts or sources for some of the datasets they've received but haven't found the time to train or integrate yet? I'm curious if any of those people or datasets would be interested in potential TalkNet models as well.
>>
>>38740212
I added the new content links since the last scrape:
- https://docs.google.com/spreadsheets/d/1T2TE3OBs681Vphfas7Jgi5rvugdH6wnXVtUVYiZyJF8/edit#gid=0

They start on row 8147. There are about 350 of them.
>>
>>38740741
>is there any potential of 15.ai sharing contacts or sources for some of the datasets
You'd probably have to email 15 directly about that.
>>
>Page 9
>>
File: 1427431.png (1.23 MB, 768x1129)
1.23 MB
1.23 MB PNG
>>38738925
The Weeknd - Save Your Tears, featuring Talknet Fluttershy
https://u.smutty.horse/mennxrpltbx.mp3

I did this song previously with both Twilight and Fluttershy singing. However, I ended up retouching the song exclusively with just Fluttershy for a friend. Again, this was just sitting on my drive. Hope you guys enjoy.
>>
>>38739460
>>38739071
Thank you again for making another one for Rainbow, I'm glad you enjoyed listening.
Actually that 40 seconds cover of Twilight is all I have from her.

>>38739114
Indeed Dash's voice can get more than often raspy for some songs. Will keep that noted with filtering, thanks!

>>38739172
>>38740390
>>38740406
Thanks you fellow anons!
>>
File: 2036109.gif (1.4 MB, 359x360)
1.4 MB
1.4 MB GIF
>>38742469
Another great quality song ponification. It's nice to hear more Flutter songs ^-^
Made another pony sync video of it, always glad to hear more mare music.

Fluttersave Your Tears:
https://u.smutty.horse/mhthilvzpmo.mp4

>>38740408
Apparently effect strips work with meta strips, so you can add an image sequence from the frames of a gif then loop it a bunch of times, make them all a meta strip then add a speed control strip then stretch/compress it and adjust until the loop is best set to the beat throughout, carefully adjusting with any part un-synced until it becomes synced again. It works great! Though it did crash a few times, so more saving is advised.
>>
>>38742648
Would be interesting if you could use a beat detection model to automatically sync a looping gif to music
>>
>>38743029
As long as the tempo track in the music creation never changed tempo.
>>
>>38743029
I agree, that'd be great and very handy. Ponies dancing to all their ponified songs. Grooving to the beats.

Seems like it'd be easy enough to do, as it's just modifying a gifs combined frame tempo to match the BPM of a given song, with delay or loop around to pair up which part of the gif will be used for the primary beat. Such a method would also definitely be more accurate than my current eyeballing method, lol.
>>
>>38743143
That being said, it's more fun to set up the image sequence for each individual beat rather than a plain loop. Especially for varying percussion tracks like I made for Luna rocking to Pendulum's Witchcraft ages ago: https://youtu.be/fRN3ji1k9-8?t=60

Though that was just 2 frames. It'd be cool if individual beats could be detected and an entire image sequence of a gif's frames could be sorted and delays adjusted to match. That way it can loop statically like I had for the Twilight Rainbow and Flutter dances, or dynamically like with the Luna clip.
>>
>>38743143
You can use this site to find the BPM of a song: https://www.all8.com/tools/bpm.htm
>>
File: 1649056618191.png (2.58 MB, 2200x2600)
2.58 MB
2.58 MB PNG
>>38742469
>>38742648
Yo this is awesome!
>>
>>38738925
>>38739460
Always nice to hear some cool RD covers.

>>38742469
And this version is nice too.
>>
>>38742648
That tiny (yay) was perfect! TN generated or pulled from the show audio?
>>
>>38742648
Thank you again for doing a Fluttershy one! ^^
Heads up, reworking an old one with Twilight! Should be finished later today.

>>38743784
>>38744589
>>38744773
Thank you! That tiny "yay" before the second verse was generated in Talknet. Voice model was set to non-singing Fluttershy and audio reference option was disabled.
>>
File: bluecute.png (127 KB, 599x1024)
127 KB
127 KB PNG
Yo Vultraz, could I trouble you for the vocal stems from Break The Limit?
>>
>>38736355
I turned your bullet points into a script here:
- https://docs.google.com/document/d/1A4kL9J_CjCM_Y1yxib5KdRF2odappM95R93-CPWeAo8/edit?usp=sharing

It's under the TalkNet section and the Image Generation section. I decided to go light on image generation, in part because I'm pretty sure I won't have time, and in part because it's hard to justify going into more detail before we start messing with these models in the PPP. Your references on it are fascinating. If you have time, can you review it? I'll be working on the slides next. Once that's done, I'll know which [examples] I can use.

Render Anon, I want to expand on the part about getting the Edge Format working. Let me know if there are any points you want me to include. Same as >>38722461. Right now, I'm going by your Edge Format document, your code, and your posts. Any additional background information you have on what it took to create your edge.py would help.
>>
>>38720786
>So should we train South Park too?
Funnily enough, I actually wanted to try that since I recently downloaded the series in 5.1 Dolby audio. But my poor computer doesn't have a CPU powerful enough to run TalkNet, so there goes that plan unless I can figure out a way to get it working without having to buy a new computer.
>>
>>38745585
(For Render Anon) I realize that >>38722461 is too broad since I have most of that already. The main thing I'm looking for is the process you went through to figure out how to convert DOMShapes to SVG. I know from the code that it involves at least these steps:
- Parsing the edge format (done from the stackoverflow post).
- Coming up with the point list approach.
- Converting edge format point lists to SVG point lists. (Maybe it's obvious from reading the SVG spec.)
- Converting stroked paths to SVG. (Maybe it's obvious from reading the SVG spec.)
- Creating closed & filled shapes from SVG point lists. (I don't know how you figured this one out. Maybe you're just better at visualizing these things than me.)

Also, why did you decide to convert to SVG? Was it because we had SVG dumps already, or did you know ahead of time that SVG was a convenient format to work with?
>>
>>38742469
The Weeknd - Heartless, featuring Talknet Twilight Sparkle (2:07, 85 BPM)
https://u.smutty.horse/mhtskdjxqbh.mp3

Twilight might of sound weird some of the verses. Generated output came out an octave higher, ended increasing the pitch by one semitone to render, lowered it to the original pitch, and tried my best preserve to matching quality.

Provided quality samples with one of the two verses in question, for those who might be interested
https://u.smutty.horse/mhtsjncuann.mp3
1)Unmodified input, generated at a higher octave
1)Changed input pitch, increased by 1 semitone via Talknet (closer to desired pitch)
2)Lowered pitch, decreased by 1 semitone via DAW
3)Lowered pitch, with modified formants
4)Lowered pitch, with Reduce metallic noise on

Image was generated with DALL·E mini, to match the original song's music video theme. Enjoy the purple drugged up rapping unicorn!
>>
File: file.png (29 KB, 229x217)
29 KB
29 KB PNG
>>38745756
Can't speak for Render Anon but my main approach to figuring out how the XFL format worked was by, making a change. Saving a copy. And checking the difference in the XFL file.
We're not the first ones to do this.
SasQ, also a brony? (His avatar seems to suggest so anyways) Wrote a conversion tool in Ruby almost a decade ago.
https://github.com/SasQ/SavageFlask ( I used this as a starting reference + stackoverflow post )

>Also, why did you decide to convert to SVG? Was it because we had SVG dumps already, or did you know ahead of time that SVG was a convenient format to work with?
My best guess is.
- SVGs are the de facto (open) standard when it comes to Vector graphics formats. Making it easier for conversion.
- The SVG format basically already includes a lot of what we need.
Such as curved primitives, (bezier/cubic curves)/fills, it's basically a super set of everything Adobe Animate supports rendering wise.
- Adobe Animate can export to SVGs as well, hence there is a guarantee we can support SVGs directly.
- We don't need to write a render engine.
- There was a paper (https://arxiv.org/pdf/2007.11301.pdf) published which used SVG for machine learning. Which can be used as baseline for trying out machine learning on leaked assets.

>Was it because we had SVG dumps already
I'm not sure which SVG dump you mean but I must note that Adobe Animate's SVG exporter has some information loss, namely the strokes are merely defined by their outlines.

> Pic related (https://github.com/SasQ/SavageFlask/blob/master/doc/FlashEdges.png)
SasQ made a visualization on how fills worked.
>>
File: 1391629.gif (696 KB, 460x460)
696 KB
696 KB GIF
>>38745768
For some reason this was harder to sync up than normal.
As someone who has also struggled to get Talknet to cooperate with certain songs, I respect the extra effort to make a song work, especially if it eventually leads to great results like this one.

Purpl - Smartless
https://u.smutty.horse/mhttairptnb.mp4
>>
File: anon 37.jpg (40 KB, 800x839)
40 KB
40 KB JPG
so me and a friend are doing a cover of tubby wubby pony waifu and are planning on posting it by tomorrow or at the very least within the week. we've made some of the dialogue overly cringe just for you guys. hope you faggots are interested. if anyone would join in for a premier. we may do that. heres a link to our channel https://www.youtube.com/channel/UCVUsDd5eAY4rCSq_66uLUqg
>>
Looks like Dall-e mini is called Craiyon now after Open AI requested a name change to avoid confusion after it's viral success.
It's also in the process of migrating to it's own site. It works already, but I guess it'll no longer work on hugging face soon is what they're implying.
>https://www.craiyon.com/
On the plus side, with two sites active, you could pretty much run an instance for the same prompt for each, effectively getting double results for the same speed. So generate at double speed whilst you can, more generations means more likelihood of decent results.

I'm wondering if this change should be reflected in the panel presentation. Like, should mentions of dalle mini be modified to reflect this? Or perhaps more simply/lazily, mention it's name change, but as it was called Dalle mini at the time of the generated uses presented, we'll be referring to that name for the remainder of the panel.
>>
>>38745882
That's all good information. Thank you.
Are you still trying to import XFL into Blender? Assuming the Blender format isn't too difficult to work with, we can probably do it now. After the presentation, I can show you how to add your own file format to my codebase as an export option. We'll need to refactor some of Render Anon's code to support non-SVG shape formats, but that shouldn't be too difficult.
If you're up for it, then any tools we develop will automatically support exporting to Blender formats.
>https://github.com/SasQ/SavageFlask
Ah right, I remember that now. I found your document again too. With that plus Render Anon's old responses to your document, I think I have all of the pieces needed to answer my questions. There are still some things I don't understand about converted Edge Format shapes to SVG shapes, but I might just need to draw out some examples of that to see how it works.

I wrote out the script here under "Parsing DOMShapes":
- https://docs.google.com/document/d/1A4kL9J_CjCM_Y1yxib5KdRF2odappM95R93-CPWeAo8/edit#heading=h.vm6o7wn8p95b
Let me know if you have any comments, and especially let me know if I got anything wrong.
>>
File: 6k8um9.jpg (71 KB, 499x601)
71 KB
71 KB JPG
>>38745901
Thank you again for another dancing pony, Hazy! Actually you gave me an idea of creating a Talknet cover parody.

>>38745768
I now present Twilight rapping about (your) porn addiction.
Talknet Twilight Sparkle - Smartless (Parody of The Weeknd's Heartless)
https://u.smutty.horse/mhtufpdnnra.mp3
>>
>>38746184
>Are you still trying to import XFL into Blender?
That project is still on hiatus. Though I've been following the threads & Cytube.
I'm currently more interested in making the leaked assets into a proper dataset that anyone can download.

>Let me know if you have any comments, and especially let me know if I got anything wrong.
Will do. Do note I'll be busy over the coming 2 to 3 weeks.
>>
>>38746330
>Will do. Do note I'll be busy over the coming 2 to 3 weeks.
No worries. If you have thoughts on it even after the presentation, I'm still curious, so feel free to leave comments on the doc at any time. I'd like to get the history right.
>>
Hey! Noobfag here, first of all pardon my french, my english is not very good. I just recently found out about this and I've been trying to train models myself using the TalkNet Colab, but I can't figure out how.
>little to none coding background btw
On step 3 of the Colab it says that I have to put training/validation .txt files on a certain path.
I have a couple of questions here:
1 - Is this supposed to be one single file with all the audio file names or multiple .txt for each audio?
2 - How do I make the validation files? Are they the same as the training files?

I found a single .txt file with the name of each .flac file on the Mega with the master file, but when I went to the Drive (because mega was really slow) it had a different type of file called .txz which contained a .json file.

Sorry about the noob questions, I'm starting out and really giving my best here. Hope I can contribute to the community as much as all the anons already did!
>>
>>38738925
seems to struggle with the really high and low notes, but the main refrain is spot on
personally i kind of hope that once perfected that this one goes viral like 'you're gonna go far kid' did 10 years ago
>>
>>38739719
Live in ~1 hour. Final assembly of script, slides and rehersals.
cytu *dot* be/r/PonyPreservationProject
>>
>>38745768
>>38746260
That TwAI voice is great. And fantastic parody man, that's top notch stuff.
>>
>>38746742
That step seems to trip everyone up, lol. If someone has more training experience then feel free to jump in, but from what I remember the training file should be a txt with the names of ~90% of your audio files and the validation should have the other 10%. The training files are used to teach the AI what it should sound like and the validation ones help you see how good a job it does with lines it hasn't heard before. So if they overlap then your "validation loss" metric will be meaningless since it will just copy its training data instead of generating something new.
>>
>>38746742
get your wav files, make sure they don't use weird characters, make a single .txt called training, list them in the format;

wavs/soundfssdfsf.wav|Gee, Bill, how come your mom lets you twinkie house?
wavs/soundfssdasddfsf.wav|Habeeb wieners.
wavs/3423.wav|Words, words, words.

until you've gone through all of your clips. Then, cut a few and put them in a new .txt file called validation. There should be no empty lines in either text file, even at the start or end. Validation should contain clips with a wide phonetic range. Training should contain most of your data (but not what you put in the validation data) to learn how to get there.

When you're done transcribing, throw all of your wav files into a zip called wavs.zip.
>>
>>38747474
>>38747835
Thanks you guys! That was really helpful, for real!
Going to get my hands on the work right now.
Usually how many files you use for a good model?
I'm thinking only using non-noisy wavs since they will provide a better outcome (I assume).
>>
>>38747918
From what I've heard, you can still use moderately noisy samples for the training data, but just don't use any noisy ones for validation data.

Of course I could be mistaken, so Anons with more training experience, feel free to confirm or correct my statement.
>>
>>38718601
>>38747188
General summary of today's PPP panel discussion:
Less than a week to go now and everyone's working hard to get everything finished. We did an initial run through of the script and discovered that we'll almost certainly end up needing an extra hour or two to cover everything we want to. There is a two hour free slot before us so sorting this shouldn't be an issue, will update as that develops.

Planning docs:
https://docs.google.com/document/d/1_kOLXFhHNF0vuiKbJXLd82YbpvLn0LDel4uG429cyis
https://docs.google.com/document/d/1NJz3rpw7Amxd_B38_o5oJdV0TIUezXlrNx8Hwy7k5eE

Slides:
https://docs.google.com/presentation/d/1udxmC5cqQejEfygn2Uxs7D4EPwwQD74JnrhYSjP1pS0
https://docs.google.com/presentation/d/1uXbQUTnm8YvNhziLPOdiyDuceFg_hAFvrrSmHAGHySE

Today's stream:
https://youtu.be/TL3hm1DpT5g
https://youtu.be/o5zmTc6jPzE

Full playlist of all streams:
https://www.youtube.com/playlist?list=PLX9nDSq9VgBMeIi8jzTrJGFttLZH_1rOO
>>
>>38747918
>Usually how many files you use for a good model?
for Talknet I see that models with 30 minutes or more sound pretty good, between than and the 20 minutes can sound still ok but not as great, anything below that like 15 minutes can still somewhat sound like the source character but it clearly will have trashy computer effect behind all the generated audio (hard to describe it other than, well, weird computer feel to it).
>>
>>38746260
Wasn't expecting my purple smart pun to be flipped and turned into it's own parody version of the song it was based on, lol. Nice work!

I've been meaning to make my own talknet parodies (or even just more songs), but I haven't gone near Talknet's interface in months now, with the last attempts turning out either crap or only partly finished. I'll just have to work through that block and see how I go.
>>
>>38745756
>Also, why did you decide to convert to SVG?
everything blender anon said. also, I don't think Animate has any other good options for exporting vector graphics. the HTML option converts every shape into JS that draws the shape with the Canvas API, which is not useful. exporting to SWF would mean that we could only view/edit the files in Animate or JPEXS decompiler (harder than using Inkscape). there may be more options I'm forgetting
Animate's SVG export isn't always guaranteed to produce identical output to Animate (see the filter effects section: https://helpx.adobe.com/animate/using/exporting-svg-format.html), but it's close enough

>Coming up with the point list approach.
honestly, point lists are an implementation detail. the goal is just to reverse edges. you could do that by directly reversing SVG paths (e.g. moveto 0 0 lineto 1 1 -> moveto 1 1 lineto 0 0), but I thought point lists made the code cleaner

from this section in the script:
>"[I] can't believe that this method was quicker to implement and required less code than texture atlases. could have saved a ton of time if I hadn't been intimidated by the edge format"
to add on, I think it took so long to get an edge format renderer because:
- intimidation/thinking there was a better way: texture atlases seemed easier to work with. we didn't know they were a dead end
- confusion: I misunderstood the edge -> shape algorithm, which wasted a lot of time
- bugs: mistakes made the initial output look unpromising, so I put it on hold for months
so there was some intimidation, but also other factors. it's not necessary to explain so much in the presentation, but it might be helpful as context. regardless, I agree with the conclusion to not shy away from reverse engineering
>>
File: 1624632197944.gif (3.39 MB, 668x534)
3.39 MB
3.39 MB GIF
>>
[WIP]
Rapemares - Spaceballs Parody - ft. Fluttershy and Applejack
https://u.smutty.horse/mhuensvmuwo.mp3

[Lyrics for completed segments]
>If you're living in a basement and you haven't got a mare
>We don't mind the sad arrangement we're still gonna rape you there
>'Cause what you got is what we need and all we do is dirty deeds
>We're the Rapemares! Oh fuck, we're the Rapemares.
>We're the masters of rape.
>(Ay, come mess around)
>With the Rapemares!
>Ugh~

>Now take off your pants anon we're gonna milk you 'til you're dry [Failed convert]
>We're destroy your little penis and it's hotter if you cry [Semi-failed Pacing]
>We'll all take your virginity, I'll start with you you start with me [Semi-failed Pacing]
>We're the Rapemares! Oh fuck, we're the rape mares.
>We're the masters of rape. [Semi-failed Pacing]
>(Come mess around)
>With the Rapemares!
>Oh fuck!
>>
File: Proud Mother Upscaled.png (16 KB, 424x478)
16 KB
16 KB PNG
Sorry I haven't contributed anything to the thread in recent months. Been drawing here and there for the aggie thread and some of the G5 threads. I've had an idea in the back of my head for a little while for a pony Twilight audio, since I haven't done her yet (among other pones). It would be a basic SoL audio (debating if lewd or not), but I need to make sure there's plenty to listen to without it being too expository or too fast.

Oh, and if Windy Whistles ever gets added to TalkNet or 15ai, I'm totally doing an audio of her too.
>>
>>38749458
That is very clearly a pony, you silly potato fucker.
>>
>>38749572
Seethe as long as your frail body will allow you, because I'm not stopping. You can bet your goddamn taint I will be making audios of G5 when those voices are available, and I'm absolutely going to be doing more EQG audios when I feel like it. And I'm gonna have fun doing it. <3
But hey, while you're here, at least give me some G4 prompts.
>>
>>38747188
Live in 30 minutes. Final assembly of script, slides and rehersals.
cytu *dot* be/r/PonyPreservationProject
>>
>>38749606
Don't forget to announce redub 4 after Q&A.
>>
>>38748706
Nice. I'll add in the information on Animate's other options being less appealing. I might add in that texture atlases seemed easier to work with. I'll leave out the other roadblocks to rendering the edge format, but it's good to know.
>honestly, point lists are an implementation detail
Fair point. I thought it was very clever, though it does sound strange next to everything else. I removed that bit.

>>38749448
>>38749593
>prompts
How are you at writing poetry? The ponies don't recite enough of it, especially Twilight.
>>
>>38738925
I appreciate you using my art for this one. Im flattered you liked it
>>
>>38749606
Live again now. Continuing final assembly of script, slides and rehearsals.
cytu *dot* be/r/PonyPreservationProject
>>
>>38749650
I haven't done poetry in a while, but I could probably do a couple poems and have Twilight recite them. Not sure how it would turn out, but might be fun!
>>
File: sweetmotherofgod.png (705 KB, 756x819)
705 KB
705 KB PNG
is there an AI image generator that can reliably generate ponies without creating abominations like these?
>>
>>38751435
https://thisponydoesnotexist.net/
>>
>>38751458
That's prompt-less.
>>
>>38751562
You are correct.
>>
>>38751435
Wasn't this shit caused because these things are also fed anthro data?
>>
>>38751435
There is not, which is why it's up to us to work towards creating one. Infinite pony via technology is possible with the right foundation. Such as well made pony specific datasets and the right AI to pair/train it with.

Even if that's just fine-tuning an existing model, which I feel would be better as it would still allow most other prompts or environmental insertion of the ponies.

Dalle Mini/Mega is still rather small, so I'd still prefer something better, however, smolness could mean less training time, allowing for better and efficient dataset testing before moving on to a better AI. Which given it's recent popularity, more and more will pop up over time. Hopefully one with great quality and adequate documentation to help us to understand how exactly to train/fine-tune it.
>>
>>38752458
Correct. These are trained GANs that give back what was put in. Big mega-GANs like DALL-E are loaded with um-curated data thus the Halloween results.
>>
>>38718601
>>38750678
General summary of today's PPP panel discussion:
The script continues to be refined and more slides take their final shape.
I've requested that the start time of the panel be moved back two hours earlier to give us more time to get through everything without rushing and leave ample time for Q&A >>38751796. Time is something that caught us out last year and there's a lot of material that we want to cover, so I'd like to make sure that time doesn't become an issue again.

Planning docs:
https://docs.google.com/document/d/1_kOLXFhHNF0vuiKbJXLd82YbpvLn0LDel4uG429cyis
https://docs.google.com/document/d/1NJz3rpw7Amxd_B38_o5oJdV0TIUezXlrNx8Hwy7k5eE

Slides:
https://docs.google.com/presentation/d/1udxmC5cqQejEfygn2Uxs7D4EPwwQD74JnrhYSjP1pS0
https://docs.google.com/presentation/d/1uXbQUTnm8YvNhziLPOdiyDuceFg_hAFvrrSmHAGHySE

Today's stream:
https://youtu.be/CVrlZulLoyg
https://youtu.be/HBB0XQPXuAs

Full playlist of all streams:
https://www.youtube.com/playlist?list=PLX9nDSq9VgBMeIi8jzTrJGFttLZH_1rOO
>>
File: 12388743123123.png (71 KB, 1415x536)
71 KB
71 KB PNG
>>38747957
>>38748180
Thanks a lot for the answers, it got me through!

I stumbled into an error during step 7 which says:

>TypeError: __init__() got an unexpected keyword argument 'period'

I also got some other error on a previous step but I managed to solve it by upgrading pytorch-lightning==1.4.4 to 1.5.10 with pip install pytorch-lightning==1.5.10

Couldn't solve the error on Step 7 tho, anyone knows how I can get past this?
>>
>>38752895
Can you expand the stack trace by clicking the up-down error next to "3 frames" (on the long line) and post a new screenshot?
Also, what was the error from before? Upgrading could fix things, but it could also break things.
>>
File: 654211231.png (116 KB, 1384x839)
116 KB
116 KB PNG
>>38752939

the error from before was
>ImportError: cannot import name 'get_num_classes' from 'torchmetrics.utilities.data' (/usr/local/lib/python3.7/dist-packages/torchmetrics/utilities/data.py) site:stackoverflow.com

Here's the new screeshot
>>
>>38752992
If you downgrade pytorch-lightning (or start over with a completely fresh runtime), then install torchmetrics==0.7.3, the problem should be fixed. The error in step 7 should hopefully also be fixed since lightning wasn't upgraded.
>>
File: sketch-1655857675332~3.png (773 KB, 1080x1084)
773 KB
773 KB PNG
Dall-e 2 experiments with ponies and has genetically altered them to be more attractive to local Moth Ponies.
>>
>>38753260
Did you get access? If so, have you been able to get any good pony art? These are good but so similar to actual product photos that they look a bit cheap/bootleg
>>
>>38753276
Unfortunately no, these results were from a YouTube video (https://youtu.be/OGbyrylx670) on different lamp interpretations via Dalle 2, this being one of them, or 5 rather.

If I got access, y'all'd likely be the first to know, lol. It'd definitely be Fluttershy in a scenic environment, like laying in an open grassy field or in a forest with sunlight shining down through the leafy forest canopy.
>>
>>38753372
Ah, that's too bad.
>If I got access, y'all'd likely be the first to know, lol.
I figured, but wanted to check just in case.
>>
>>38745902
uploaded. hope you fags like it
https://www.youtube.com/watch?v=Llk77Z9naFI
>>
>>38753620
This is both awesome and adorable. Nice work ^-^
>>
>>38753643
thanks, hazy. was totally expecting insults.
>>
>>38753260
kinda like that heart cadance lamp.
>>
>>38753645
I unironically like that song.
Partly because it's cute, and partly because I really like Greg's RD voice throughout the MAS and RDP series.

It's only a matter of time before there's a canon-pony redub version for all fan impression works

>>38753653
Same here, I think it could actually work well. Just needs a more more R&D to flesh out the correct color scheme and fix other issues. I'd happily buy and/or make one after that.
>>
In case any of us get Dalle 2 access. Here's a list of 100 ideas I just came up with for it that I'd love to see.
https://u.smutty.horse/mhuqhckqveo.txt

Don't have to use any of them of course, but I hope they were interesting or amusing. I also encourage you to compile your own lists if you're signed up for the waitlist, just in case you get lucky and/or are unsure of what to generate next when you do.
>>
When training a model for Talknet, does the volume need to be consistent? What about quality?
>>
Spoilers for the panel.
>>38752559
Animation samples for where we want to show the current state of animation data: https://drive.google.com/file/d/1YKMjr6SmZcMMNedT2L5edvsiWCaN9Kar/view?usp=sharing
These are all samples that you labeled as "unclean". I figured they'd make for more interesting animations than the clean ones. I can do the same with clean samples.
>>
I kind of want to make a short story like Tax Breaks. Who wrote it?
>>
>>38754219
https://www.fimfiction.net/story/185725
Its literally in the video description
>>
>>38754048
The volume of each clip doesn't have to be exactly the same (e.g. it makes sense that a shouting clip is louder than a whispering clip), but you don't want extremely quiet or loud clips.
For quality, higher is always better, but if you have limited high-quality data, it's fine to use some noisy data. Exactly how much to use probably depends on your specific case.
>>
>>38750678
Live in ~20 minutes. Final assembly of script, slides and rehearsals.
cytu *dot* be/r/PonyPreservationProject
>>
File: 98123784123.png (92 KB, 1468x833)
92 KB
92 KB PNG
>>38753208
thanks anon, this solved it yesterday. But I got another error today.
I left the Step 11 running while I was asleep but it ran out of GPU time, so I subscribed to Colab Pro and tried to re run again. It got an error that I unfortunately didn't screenshoted.
I tried to solve it by cleaning the trash and running step by step again (including install torchmetrics==0.7.3) and now, on Step 5, there's a new error.
Am I being unlucky, dumb or is this normal?
>>
>Page 9
>>
>>38755326
General summary of today's PPP panel discussion:
Everyone's working hard to get slides and script finalised; we're almost there.
Also did some more rehearsals.

Planning docs:
https://docs.google.com/document/d/1_kOLXFhHNF0vuiKbJXLd82YbpvLn0LDel4uG429cyis
https://docs.google.com/document/d/1NJz3rpw7Amxd_B38_o5oJdV0TIUezXlrNx8Hwy7k5eE

Slides:
https://docs.google.com/presentation/d/1udxmC5cqQejEfygn2Uxs7D4EPwwQD74JnrhYSjP1pS0
https://docs.google.com/presentation/d/1uXbQUTnm8YvNhziLPOdiyDuceFg_hAFvrrSmHAGHySE

Today's stream:
https://youtu.be/KAdfe5GqEp4

Full playlist of all streams:
https://www.youtube.com/playlist?list=PLX9nDSq9VgBMeIi8jzTrJGFttLZH_1rOO
>>
I don't really understand how I'm supposed to use the offline talknet training. I put the required materials into my google drive and tried to path to them with the notebook, but it doesn't work. Am I supposed to modify the notebook so that it works with my google drive?
>>
File: sketch-1655951038058.png (79 KB, 1244x1080)
79 KB
79 KB PNG
>>38757153
I've personally never trained with TalkNet specifically, but when I was training with TT2 I linked the Colab to my Google drive, located the folder containing all the training data, and moved the folder to the location the Colab script was looking for the files by default. Renaming it afterwards if necessary.

You can link your Google drive by clicking on the folder icon on the left, pressing the folder with the Google drive icon, then running the code box it creates.

Unsure if this is in any way beneficial, but hope it helps somehow.
>>
>>38757250
To be honest, I've just been stumbling around trying to figure out how the hell to train my own model. I only ever vaguely found that talknet training thing here: https://github.com/SortAnon/ControllableTalkNet/blob/main/TalkNet_Training_Offline.ipynb

Does anyone know if there an actual tutorial for doing this somewhere? I've pretty much just been following via what >>38746742 asked, but I have gotten stuck getting the actual training process going.
>>
>>38757153
which colab are you using?
>>
>>38757406
That's probably the problem here. I attempted to use the ipynb I linked in my previous post in Google Collab.
>>
File: file.png (13 KB, 739x174)
13 KB
13 KB PNG
Wait what happened to the doc?
>>
>>38757410
this is the one that's linked in the documents
https://colab.research.google.com/drive/1Nb8TWjUBJIVg7QtIazMl64PAY4-QznzI?usp=sharing
this is the one you're supposed to use
>>
>>38757424
Thanks, that looks far better so I'll give it a shot. Been struggling to find anything with the main doc down.
>>
>>38757421
Apparently spam? Which seems odd.

I'm still suspicious this could've been intentionally spammed in order to block it just to spite us or something. It could be some other reason, but that seems to be the most likely potential reason and motive to me. There's still the backup doc though.
>>
>>38757421
>>38757433
I wonder if Uberfuck had something to do with this.
>>
>>38757426
There are backups of the most recent version (from >>38683242)
https://u.smutty.horse/mhmqxnbjaye.docx
https://u.smutty.horse/mhmqxnefvuj.pdf
>>
>>38758019
I wouldn't be surprised, honestly.
>>
>>38756391
It's a problem that has occurred before: https://desuarchive.org/mlp/search/text/numpy.ndarray%20size%20changed
Try running `pip install -U numpy` after Step 4. The version installed should be at least 1.20.0.
For more detail, see https://stackoverflow.com/questions/66060487/valueerror-numpy-ndarray-size-changed-may-indicate-binary-incompatibility-exp
>>
File: 837316231.png (92 KB, 1522x742)
92 KB
92 KB PNG
>>38758197
I see, I managed to get it working after just reseting the computer and been inactive for a while.

I reached another problem tho, I upgraded to Colab Pro so I had more time to run the hifi-gan but it reached it's capacity while I was asleep. Since it already ran enough time, I wanted to just package the model, when I tried, it gave me an error
>>
Any advice on training for a small dataset (e.g. 4 minutes)? I know there's not really anything to do about it, but I was wondering how to handle the validation list. There's already barely any data to use in the first place.
>>
File: 3 anime face sliders.webm (2.94 MB, 926x720)
2.94 MB
2.94 MB WEBM
https://twitter.com/t_takasaka/status/1479784432697749513
some interesting thing sniped from /3/, it dosent seems that the code is open but it's still, it is something interesting to look at.
>>
>>38758464
Just run `import os` in any cell and try again
>>
File: 78787321.png (30 KB, 766x354)
30 KB
30 KB PNG
>>38758508
it worked but now, another error. Sorry for bothering so much
>>
>>38758539
I think it worked, I just had to mount the drive again
>>
>>38758541
Glad you got it working, after all that effort. Hope your model ended up okay and look forward to hearing it.

Trial and error; it works eventually.
>>
>>38758547
Just a final bump on the road, when I tried running the custom model, it says that the model failed to download even though I made it accessible to anyone with the link

And thanks so much for all the help so far, couldn't make it without you anons
>>
>>38758507
Freaky. This like auto-tune for drawings, but cranked to 11.
There is finally hope for my artistic career!
>>
>>38738925
How did you make it work with overlapping voices?
>>
>>38755326
Live in ~30 minutes. Final assembly of script, slides and rehearsals.
cytu *dot* be/r/PonyPreservationProject
>>
>>38759020
I would imagine he sing the first part and than song the second part and then combined them together in audacity.
>>
>Page 9
>>
>>38759317
Cytube down?
>>
>>38760825
It's down for me too, so yes.
>>
>>38760841
Can you post the youtube stream link?
Imagine if this happened tomorrow, when we're live.
>>
>>38760852
Nvm it's back up.
>>
>>38760852
Man, don't jinx it like that. That'd suck.
>>
What was that playback and transcription software I saw posted here a few months ago?

One of the talknet models I made earlier was terrible, because at some point, the training.txt list differed from what my media player thought was the proper order.
>>
File: form but wrong horse.png (1.2 MB, 1688x945)
1.2 MB
1.2 MB PNG
>>38753276
not that anon but,
>I'm in.
Unfortunately Dall-E has a bias against preexisting characters and trademarks, pic related
>>
>>38761188
Nice!
>bias against preexisting characters and trademarks
Still, it looks good! Compared to DALL-E mini/Craiyon, it's clearly recognizable as horse art, with defined lines and textures, not blurry blobs. And compared to TPDNE, it can generate multiple poses and full-body images, not just faces.
Does it fail to generate any MLP characters? I would expect it to know about the M6, at least.
>>
>>38761188
>Unfortunately Dall-E has a bias against preexisting characters and trademarks
It seems to make a very good Homer Simpson though.
https://dalle2.app/?search=homer
>>
File: shipping.png (1.25 MB, 1690x932)
1.25 MB
1.25 MB PNG
>>38761218
they're all still off
also I'm out of attempts for the day so I'll have to wait 19 hours for them to refresh
>>
>>38761285
Still pretty good, you could pass it off as someone's OC. Thanks for trying it out.
>>38759317
Might be worth sticking these DALL-E 2 samples into the slides as a last-minute addition?
>>
>>38761188
>>38761285
Holy shit, nice!
I envy you a ton, Anon.

So it's limited daily generations huh? I wondered about that. How many do you get daily? Also, how long ago did you apply for it?

>>38761439
I definitely agree, especially the two Twilight-like ponies in front of the moon (top left of second batch)
>>
File: air force dash.png (1013 KB, 1686x948)
1013 KB
1013 KB PNG
>>38761698
50 generations daily, each generation does 6 images
here's a past one I did, it's quite interesting to note that it recognizes that "Rainbow Dash" implies a pegasus, but it seems to filter out anything that actually looks like RD, likely due to some sort of built-in anti-copyright rules
>>
>>38761698
>how long ago did you apply for it
beginning of April, got into it maybe a week ago
has rules, though
>no hateful imagery or speech
>no harassment
>no self-harm
>no violence
>no lewd
>no shock content
>nothing illegal
>no politics
>no real people
>you need to own the rights to the image you're uploading or have permission to use it

What it can do:
>specify what kind of art style and medium
>create variations of either an image you uploaded or that it generated
>edit an image you upload (not sure how this works, only tried it a handful of times)
>>
>>38761810
We need to violently attack copyright before its too late
The time is now
>>
>>38718601
Instructions for downloading animation data (render traces), /mlp/ images, and derpiboru images:
- https://ponepaste.org/7515

Downloads for /mlp/ comments and metadata & derpibooru image metadata:
- https://drive.google.com/drive/u/2/folders/1v-qOV0jUKNKdyxJzunHehMMMeln1d36f
These are parquet files. In python, you can use the pandas library to load them using:
>import pandas
>dataframe = pandas.read_parquet('/path/to/file')

Everything is up-to-date as of a few days ago. I have a bit more metadata that I haven't uploaded yet for linking images together across datasets. I need to update it with information from my latest scrape. I'll post that after the panel. I also have /mlp/ image icons, which I haven't uploaded yet. I'm running low on server space, so I might need to add another disk before uploading that.
>>
File: star.gif (1.63 MB, 600x600)
1.63 MB
1.63 MB GIF
>>38762176
I saw the note in the slides about fuzzy/grainy gifs. you can use https://gif.ski/ to generate high-quality gifs, see pic related. it takes a video or a series of PNGs as input
>>
>>38758468
The main purpose of validation data is to tell when you're overfitting (when training loss is going down, but validation loss is going up). With so little data, you don't want to put any in validation, but you still need to know how long you should train to avoid overfitting. Maybe try this:
- Edit the cell's code and set "cfg.trainer.check_val_every_n_epoch" equal to 1. This prints validation loss for every epoch, not every 5 epochs. It will be slower, but your dataset is small, so it's fine.
- Now take a little bit of your data (maybe 20s to 45s?) and put it in the validation list. Don't let it overlap with the training list.
- Start training. Watch the validation loss. It should drop, then plateau, then start to rise. For example, you might get: 1.5, 1.1, 0.9, 0.8, 0.75, 0.73, 0.72, 0.72, 0.71, 0.8, 0.9, 1.1 (these numbers are completely made up). I would say it drops from 1.5 to 0.73, then plateaus around 0.72-0.71, then rises from 0.8 to 1.1.
- Find the number of the epoch where it starts to plateau. In this case, if 1.5 is epoch 1, then the plateau starts at epoch 7. This is the point where the model is learning less and starting to overfit.
- Now, empty out the validation list and put everything in the train list. Change "epochs" (the max number of epochs to train) to whatever number you got for the plateau. Change "load_checkpoints" to False (we want to start fresh). Finally, train again.

The idea here is to first use validation data to see how long it takes for the model to overfit. Once we have an number, we then train with no validation data and stop at that number of epochs--hopefully the exact point where the model is starting to overfit.
It's true that the point of overfitting will change depending on whether or not you use validation data, but since you have so little data overall, I think it won't change by much.
If it's not working or overfitting too quickly, try a lower learning rate like 1e-4, 1e-5, or something in between.
I've never done this myself, so I don't know if it will work, but hopefully this helps.
>>
>>38761810
Haha, literal Blue Dog xD

Also strange how for dogfighting it's using airliner style civilian aircraft rather than anything smaller and military. I bet that likely means military craft are harder to get without mentioning specifically like "F-22 Raptor".

>50 generations per day
Woah, that's a lot more generous than I was expecting, that's awesome!

>>38761827
>Strict rules
Makes sense, definitely keeping that in mind. I wonder if a horror creature counts as "shock content".
>Image revisions/variations
Sweet, that sounds very helpful for perfecting decent results. Bummer that it likely uses another daily credit though.

If you ever happen to run out of ideas for generations and/or would like some inspiration for prompts, feel free to try out or create variations of my 100 ideas too try out I mentioned in >>3875399, I'd love to see how any of those turn out. Though as you can imagine, the pony ones would be more relevant here. The rest (or otherwise) I'll do when my time comes.
>>
>>38762513
The ones I mentioned in >>38753991*
Looks like I cut off the 1.
>>
File: thumb.png (213 KB, 666x558)
213 KB
213 KB PNG
https://youtu.be/3UUmJNyZOyo
Another Dashie racing song. My first real attempt at eurobeat.
>>
>>38762528
Saw this within the first minute of upload, lol.
Another great song once again ^-^

Synth backtrack reminds me of F-777 a little.
>>
>>38762528
Might put this in my playlist. I only did that for Not Okay and Welcome To The Village (just my particular tastes). Good job.
>>
>>38761188
>>38761285
>>38761810
Thanks, I've added these to the slides.
>>
>>38753991
>>38762522
will do, I should be able to get through the list in a few days, focusing on ponies first, dependent on if anyone other requests
>>
bumpo
>>
File: 1648477494603.gif (1.9 MB, 1000x1000)
1.9 MB
1.9 MB GIF
>>38762528
>>
File: Schedule.png (2.61 MB, 1256x6700)
2.61 MB
2.61 MB PNG
Here's the planned schedule for the topics that will be covered in the PPP panel tomorrow. It’s a long panel with lots of different topics so this will help (You) get an idea of what’s happening and when. Blue sections are non-technical, orange is somewhat technical, and yellow is VERY technical. Don’t be intimidated by the apparent large number of technical segments, it’s just been segmented to show all the topics that’ll be covered. Overall, the panel will be ~50:50 technical/non-technical for its run time.

As an overview - Synthbot, GothicAnon and I will share a general history and status update, followed by in-depth technical explanations of the new tools and datasets and a look to the future of AI on /mlp/, concluding in live Q&A. We also have, for the very first time, a guest appearance by our very own robo pony TwAIlight Sparkle!

Since it’s a live event, the timings may vary slightly from what’s written on the schedule, it should be used as a rough guide only.
>>
>>38762310
Thanks! I re-rendered the ones we're using for the presentation.
>>
File: 1647276479711.png (829 KB, 3700x2750)
829 KB
829 KB PNG
>>
File: whats a fluttershy.png (1.81 MB, 1684x937)
1.81 MB
1.81 MB PNG
>>38762522
case in point with the anti-existing bias:
>What's a Fluttershy?
>>
File: pony text.png (821 KB, 1683x932)
821 KB
821 KB PNG
>>38762522
Even being completely clear with what you want produces chinese knockoffs
>>
>>38765063
Here's a full size of the top middle since that seems to be the one with the best composition
at best it seems to be something you can use to create a springboard for AI-generated OCs
>>
>>38765050
That's a little disappointing. I expected it'd still have her in it, as other AI were able to do it. Maybe that particularly prompt was to specific, focusing more on the forest and the light rays?
>>38765063
>Chinese Knockoffs
This is to be expected, as it's more so trying to get the style. I think it did very well in that part.
>>38765072
Oh my gosh, she's adorable!
Now that's a quality pony right there, good proportions, nice posing, looking at you with a smile, nice mane shading. Only a couple errors with the front legs, but easily fixed with a little editing.

We should give her a name, maybe something using the words/letters it generated. Like... "Litt" something, or including "Mangic", "Little Mango sounds the closest, but doesn't match her colour scheme.
>>
>>38765221
Littl Mangic?
Uses Litt and Mangic and hopefully clearly says how it's a "little big of manmade magic?"
>>
Hey SortAnon, I'm not sure if you're still working on speech models, but there's a ton of good feedback on TalkNet here from Vul: https://www.youtube.com/watch?v=S_QtySdMIB0
>>
File: variants.png (701 KB, 1032x694)
701 KB
701 KB PNG
>>38765221
here's the variations ability
it took that image as the input, determined what was in it, and then produced variations of what it determined to be in the image
fucking sucks that you cant use an image you down legally own and I dont want to risk getting b& from this by feeding it a twilight sparkle to create variants of
I know fanart is technically in a weird legal limbo no one actually cares about but I have no idea how fickle the devs are
>>
File: 9509ec67.png (4 KB, 271x47)
4 KB
4 KB PNG
i'm getting dalle-2 access soon as well, yay
>>
>>38765308
>Dall-E letting /mlp/ infiltrate its ranks
Equestria will grow larger
>>38762522
>pic
>>
>>38765072
this is so precious
I'm crying
>>
>>38761827
>no violence
>no lewd
how can I cope?
>>
File: dall e self sona.png (538 KB, 1699x956)
538 KB
538 KB PNG
>>38762522
I dont think the acronym works out so well, but the full title is interesting
>>
>>38765319
Ha, this is awesome.
>>
File: okay this is weird.png (1.01 MB, 1689x940)
1.01 MB
1.01 MB PNG
>>38762522
>keeps making blue horses with pink manes as in >>38765072
I think Dall-E has a preferred color scheme
Also surprised it managed to make a fairly decent Glimmy
>>
>>38765349
Try specifying their mane/coat colors. Also race.
>>
File: 1545803609300.png (207 KB, 399x339)
207 KB
207 KB PNG
>>38765340
>three-legged anthro centaur
Jesus Christ how terrifying.
>>
>>38765344
>>38765352
if you'd like the full 1024x1024 of any image just ask asap before it gets overridden, it'll store the 10 most recent generations on the sidebar
>>
>>38765359
you're waifu calls me a three-legged anthro centaur
>>
try "a little Pegasus pony with yellow fur and a pink mane"
>>
I've seen people using prompt enhancers like "3D Render, Unreal Engine, 4K Detailed" but it wouldn't probably work as well on pony.
>>
>>38765380
>a little Pegasus pony with yellow fur and a pink mane
>AI suddenly creates fluffy pony art
panic.kornheiser
>>
>>38765392
here's the top left since it actually resembles fluttershutter
>>
File: princess fucking luna.png (978 KB, 1703x943)
978 KB
978 KB PNG
>>38762522
I'm editing these suggestions since I've learned how the lexicon if the bot manages to work and how you need to try and trick it into making what you want it to make
that said this prompt is cursed
>>
>>38765285
The bottom left gives a bit more info to potential naming, as it looks similar enough to be the same but by a different... "artist". 1028x1028 of that one please ^-^. Multiple references to same mare is always good.
>Water droplets cutie mark
Perhaps she's a gardener who enjoys watering all the plants day by day.
>Purple eyes
True eye color confirmed too.

>>38765308
Hype!

>>38765340
I really like the top left and bottom right ones, you should save them both are share (via https://smutty.horse/ links). The bottom right one while messier does have a lot more style and identity to her.
>>38765349
I had a feeling the full name was required for better results. Daaamn, that bottom left one. Now that's a damn cute pony, aside from what's going on with her... hoof extensions? Top mare.
>>
>>38765394
Adorable Flutters! I'm really diggin the pencil/chalk like style. The closed eye, head down and ears back pose was a good artistic direction/choice to minimize errors.
>>38765416
Top right. Fedora wearing, sunset shimmer dyed human of dark complexion spooks a horse.
Bottom right. Human meets an abomination of a mare, offers hand in friendship anyway ^-^.
>>
>>38765422
>you should save them both are share (via https://smutty.horse/ links)
https://u.smutty.horse/mhvvwlsmzuo.png
https://u.smutty.horse/mhvvwluujlf.png
>Daaamn, that bottom left one
https://u.smutty.horse/mhvvwlutfbz.png

>>38765433
yeah, the AI has severe problems when it has to draw a character that would have two eyes showing at the same time, best solution I've found was to cover it up with sunglasses, but that's not much of an option with MLP
>>
>>38765436
>Bottom left
>https://u.smutty.horse/mhvvwlsmzuo.png
Oops, I misspoke. I meant the bottom right one, lol. Which I think is one of the best quality ones yet.
The one with the flowing purple mane, with the hoof extensions and the somewhat star-like cutie mark.
>>
>>38765444
ah no problem, has some real Lisa Frank vibes to it
https://u.smutty.horse/mhvvyayesth.png
>>
Made another attempt at Fluttershy with Dall-E
https://u.smutty.horse/mhvvymemylc.PNG
>>
>>38765462
Attempt 3 is closer
https://u.smutty.horse/mhvvzeiidef.PNG
Attempt 4 changes one word from 3
https://u.smutty.horse/mhvvzetpwel.PNG
resulting in a successful oil painting of Fluttershy
https://u.smutty.horse/mhvvzghucho.png
Horse has been successfully created
>>
>>38765481
Interesting thing of note is that I never specified it was a pegasus, but the AI added in what is clearly a wingtip anyway
We'll make a horsefucker out of it yet
>>
>>38765349
Got both middles and upper left? They look pretty
>>
>>38765522
I figured they were nice enough to save, here you go
https://u.smutty.horse/mhvwbzsrfkd.png
https://u.smutty.horse/mhvwbzuxdbk.png
https://u.smutty.horse/mhvvyayesth.png
>>
>>38765462
Still add some occasional images of results directly in thread. Just use smutty when there's like... multiple of them just to keep the bump count low over time and stuff.

More Dall-e 2 hype, and pony experimentation is welcome. Especially if that means better results for less tickets for when (at least some of) the rest of us also gains access.

So far we've learned:
- Focusing too much on environment means unlikely pony.
- Different style results from MLP:FIM and it's non-acronym name.
- Describing a canon mare can get very close results to intended mare.
- Dalle 2 is definitely not free from pony abominations
- Princess Celestia is incompatible with nebulas?

>>38765481
Love the horsey ones too.
Horsified mares are cute.
>>
>>38765572
the biggest lesson is that there's an entire lexicon you have to work with, and if you want a canon horse you're going to be actively fighting the damn AI since it doesnt want to create fanart
I'm trying to get to more show-accurate models instead of oil painting (which honestly makes nearly anything look far better) and that's having some mixed results
removed medium specification? deviantart
https://u.smutty.horse/mhvwhhhucdp.PNG
tell to to use flash? neopets
https://u.smutty.horse/mhvwhiaklpk.PNG
slap in vectors ontop of that?
https://u.smutty.horse/mhvwhhxryna.PNG
starting to get fairly close now, especially with pic related
>>
>>38765602
Rephrasing always seems to be a requirement for most AI if initial results don't show the desired outcome. I'd suggest experimenting with the lexicon and syntax. Like seeing of commas are more effective than periods for descriptive separation. If character names are more likely to generate the desired output if the name is said around the end rather than the beginning, changing the order of the subjects. If negating key words like "cartoon" helps to get less bias towards more "toony" works that aren't up to the style or quality of FiM. Identifying styles that optimize results. Stuff like that.

>https://u.smutty.horse/mhvwhhxryna.PNG
That bottom left one could be a passable Sunset.
>>
>>38765647
Went on a tangent just to see if I could get it to make fluttershy in any art style first
https://u.smutty.horse/mhvwmkfogqq.PNG
led to
https://u.smutty.horse/mhvwmkollnp.PNG
bottom left got my eye so I generated variations of that to produce better fluttershys
https://u.smutty.horse/mhvwmktxrdr.PNG
bottom middle here is weird
if Dall-E is not for lewd, then why does she have bedroom eyes?
https://u.smutty.horse/mhvwnhssjmm.png
I'll drop cartoon now and fine tune until I run out of requests for the day
I'll post anything notable tomorrow
>>
>>38765698
>https://u.smutty.horse/mhvwmkfogqq.PNG
can I have the pony on the top right?
>>
>>38765725
Yeah that was too fucking funny for me to not save
https://u.smutty.horse/mhvwpiksqqq.png
>>
Yandex just published the parameters of a 100B parameter model.
https://github.com/yandex/YaLM-100B
>>
>>38765816
>100 billion parameters.
>800 A100 graphics cards used to train.
>Requires 200GB of free disk space available before downloading weights.
>Tested running using 4 A100s.
>Requires around 200GBs vram to run.

That's interesting and massive, but I feel it's kinda seems useless for us given our limited tech.
We're lucky enough to have a 6B model up and running well, 100B sounds way out there, lol.
Still, hilarious to see how crazy big AI models are being developed these days.
>>
>>38765816
Sweet, if I ever win the lottery I'll set it up.
>>
>Pinkie loses it and decides to fix the show
https://u.smutty.horse/mhvwzshlrkb.PNG
>>
>>38765481
Is it too late to get full version of the bottom right of Attempt 3?
It's a good one, with it's very horesy snoofa and all.
>>
First time trying anything like this. I thought it was a little funny.
https://vocaroo.com/1gSDPLGVfBDr
>>
>>38765340
Aww, look at that cute little zombie on the top right. Looks like she hasn't eaten in days.
>>
File: 000664.jpg (28 KB, 618x458)
28 KB
28 KB JPG
Why the fuck does this keep happening. I hate github so much. Do we have a viable alternative?
>>
>>38766766
Gitgud.io
>>
>>38766766
Did you sign up for two accounts? That happened to me before.
>>
>>38766766
What was nuked this time?
>>
>>38767333
Going by posting time, I'm assuming Catch the Mare. >>38766878
>>
For any anons not sure where to post questions for the Q&A:
>>38767237
>>
File: Thumb.png (433 KB, 776x950)
433 KB
433 KB PNG
PPP FiM E4 AI REDUB
The time has come to continue the /mlp/ AI Redub series! This time, we're redubbing season 1 episode 4 "Applebuck Season" with AI, and we want the whole board to contribute. See the Info Doc for more details.

Rules & info:
https://docs.google.com/document/d/14YzE-WBTH4xvwP2vU-Uk28T3j-iWKhf5s3DS-LKv4M8

The spreadsheet:
https://docs.google.com/spreadsheets/d/1uWLs6z1nz49VNfBfI4rB3rKHiCj81YDO6j0U987zuHs
>>
>>38767902
Claiming 4-17
>>
>>38767902
ziggers tonguing my ponut
>>
>>38767902
claiming 4-15, AJ hard of hearing
>>
>>38767902
claiming 4-19
>>
>>38767935
also claiming 4-25 plz
>>
Don't take all the good clips, fags. I'm watching the con right now.

>>38767963
One (current) claim per anon. Maek 4-19 before claiming 4-25.
>>
>>38767980
>One (current) claim per anon
sorry about that, claim 4-25 dropped.
>>
>>38767902
Gonna give 4-14 a try
>>
>>38767902
Claiming 4-1.
>>
>>38767902
Claiming 4-8.
>>
>>38767902
Claiming 4-3
>>
File: fucking why.png (88 KB, 1917x1017)
88 KB
88 KB PNG
>>38767902
i'd like to claim 4-18
and also vent that i fucking hate audacity and seem to be the only person on the planet who can't get the damn thing to work properly
>>
Dang it. Due to having fallen asleep due to con hours for my time zone, wasn't able to claim the section I already had an idea for.

Oh well, guess I'll have to look through the sections and hope I get another idea that's comparable. After the Shawn panels though.
>>
>>38768458
You can do it as a backup.
>>
File: file.png (35 KB, 723x362)
35 KB
35 KB PNG
i was trying to get the gpt pony thing that they were using on the panel when I get to the last cell i get this error
>>
>>38765647
Can DALL-E 2 take an image generated by Craiyon and improve it?
>>
>>38766232
unfortunately it is too late, I generated a bunch of new images, trying different punctuation, orders, and such and I'm not really getting any sort of progress
I'm guessing it might be too specific a style for it to match the show, especially when it's likely going to be using search results to learn from, search results that would be filled with fanart
this is probably the cleanest one I got last night

>>38769297
possibly, though you'd have to see if you own the rights to use images from Craiyon to do so
>>
>>38769698
>Too late
Bugger, I had a feeling that might've been the case. Oh well.

>see if you own the rights to use images from Craiyon
According to their FAQ, so long as it's personal use and not commercial, they don't care.

>"Can I use the images generated through craiyon?"
>"Yes, feel free to use them as you wish for personal use, whether you want to share
>them with your friends or print on a T-shirt. For commercial use, please contact us."

So I don't think using them with Dall-e will be an issue. Unless the Dall-e 2 devs are very nitpicky and because the Craiyon output is of a character you don't have the rights to or something, it's not allowed? I honestly think the rights thing is intended so you can't CLAIM a work as your own using work that isn't. So long as you're not claiming it's your own and aren't infringing on anything, I think you should be fine.
>>
I’m really flattered that I managed to get the same song played in the mega compilation played nearly back-to-back four times, at the same time I’m a bit embarrassed by the fact that half were my trial-and-error posts from before SortAnon had to beat the proper settings into my head.
>>
>>38767902
I'm pretty much stumped on ideas the the narrative of this episode, so I'm prolly not gonna claim anything until I see some other sections done to give inspiration, or I come up with another idea over the next few days.

In the meantime, I'm gonna attempt to complete the concept for 4-1 I shared back in mid-May to at the very least use for a rejected/deleted scenes thing.

Just wish I hadn't have fallen asleep mid-stream so I could've claimed it initially, sometimes I really hate being in an opposing timezone.
>>
>>38765869
Skit-worthy green
>>
File: urzas ai.png (696 KB, 442x874)
696 KB
696 KB PNG
https://www.urzas.ai
So there is a new art ai that makes automated magic the gathering style cards, it doesn't seems to have a good grasp at the pony shape but do throws the correct colors where they should be.
>>
>>38770482
Looks like it uses wombo to make the art
>>
>>38767902
4-2 for me
>>
File: Discord_352jo33v1s.png (57 KB, 1032x370)
57 KB
57 KB PNG
novelai has a new text-to-speech thing that was trained on ponies (among others)
but they can't be accessed directly because it's supposed to be for random voices
buuuut turns out they're still sorta accessible via guessing
demos follow, see pic for seeds

twilight sparkle: https://cdn.discordapp.com/attachments/988614902111416371/989272695424315512/NovelAI_TTS-seed.menu.Dear_Princess_Celest....mp3
rainbow dash: https://cdn.discordapp.com/attachments/988614902111416371/989120368822337616/NovelAI_TTS-seed.PrincessTwilightSparkle.It_needs_to_be_about....mp3
applejack: https://cdn.discordapp.com/attachments/988614902111416371/989471487385821214/NovelAI_TTS2-seed.seedmix_nest_cadence_Aurlonhole_80.Well_howdy-doo_Mis....mp3
spike: https://cdn.discordapp.com/attachments/988614902111416371/988916946076590111/NovelAI_TTS-seed.polarized_overjoyed_echium.Spike_the_dragon._2_....mp3
sweetie belle: https://cdn.discordapp.com/attachments/988614902111416371/989632828268507136/NovelAI_TTS2-seed.sawt.Were_gonna_have_the....mp3
trixie (not pictured, seed is `lineup_enthusiastic`): https://cdn.discordapp.com/attachments/988614902111416371/990419756085571644/NovelAI_TTS2-seed.lineup_enthusiastic.You_may_have_vanquis....mp3
>>
>>38771566
for copying:
Twilight Sparkle: menu
Rainbow Dash: PrincessTwilightSparkle
Pinkie Pie: seedmix:78ty7olk'|intonation:78ty7olk'+infrequent_gathering|cadence:Claea+78ty7olk'
Fluttershy: venue_living_cougar
Applejack: seedmix:nest|cadence:Aurlon+hole_80
Apple Bloom: seedmix:camp_35-mad|intonation:x/a&;BtT{u"=D*g#zS4pVL9s8bUNJ~vK_G.6?>|cadence:f?p.)>J7#@92VKb!
Sweetie Belle: sawt
Spike: polarized_overjoyed_echium
Princess Celestia: jeep_4660_jealous_ossified_hippopotamus
Discord: maroon_maple_league_ferry_chattering
>>
>>38771566
oh god that AJ is so cursed
>>
>>38771566
Any examples with text not from the training set?
>>
>>38771566
Those sound great, except they got AJ speaking with a non-regional dialect.
>>
>>38771566
how I wish we could take a peek at the behind the scenes stuff.
>>
>>38771566
Wow, these sound just amazing!
That AJ sounds like she's Blended with Celestia though, lol.

It's great to see ponies have spread far and wide, and are steadily improving in quality, even outside thread. I look forward to trying out this Novel AI TTS solely to try out these voices and how well it compares to 15.ai's lineup.
>>
>>38771566
Is it just me or are these not very good
If these are the best ones I'm not looking forward to the actual model
>>
>>38771568
So I was cycling through random seeds with a dirty lewd prompt about how much the character wants Anon's hard dick. And suddenly one of the seeds sounded EXACTLY like Granny Smith saying it.

I laughed so fucking hard that I accidentally tapped my phone in the wrong place and it ended up reloading a presaved voiced and removed the seed. Dang it. Serves me right for trying with my phone where I can't into text boxes. Guess I'll keep trying. There's bound to be more pony-like voices amongst them, just gotta find 'em.
>>
File: picrel.png (156 KB, 2654x328)
156 KB
156 KB PNG
>>38772161
Nowadays I'm always skeptical whenever a TTS company says they have a "better model than 15.ai" if they show off samples without letting people actually use said model because comparing it to 15.ai is just the de facto baseline in TTS these days. Unless they put it up on a site where we can easily use it to make ponies say nigger, I'm gonna take it with a grain of salt.
>>
>>38772082
>Princess Applejack
>>
File: file.png (133 KB, 1848x1074)
133 KB
133 KB PNG
>>38772226
you can try it if you put in the seeds
>>
>>38767902
claiming 4-25 Stubborn Pride Aside
>>
File: 589676_edit.png (498 KB, 695x900)
498 KB
498 KB PNG
I FOUND IT! - "course_conglomeration"

Man I've been having an absolute blast with this. I've also learned a few things along the way.
If you add periods to the start of your prompt input, you can generate a different result with differing emotion or emphasis. It's very useful, though I needed minimal of that for this one.

>Granny wants you to pound her, Anon.
>https://u.smutty.horse/mhwrexkeusz.mp4
>>
>>38772568
Not raspy enough, immersion ruined, 0/10
>>
File: file.png (177 KB, 1436x432)
177 KB
177 KB PNG
>>38771566
just tested it on novelai myself, it's really clear and has great emotions when you force it but it's quality when you use long prompts drops compared to 15.

Seashore Tongue Twister 15: https://u.smutty.horse/mhwrjdromcr.wav

Seashore Tongue Twister NovelAI: https://u.smutty.horse/mhwrjdikqgs.mp3

Mares! 15:
https://u.smutty.horse/mhwrjdmtiks.wav

Mares! NovelAI:
https://u.smutty.horse/mhwrjdhkpki.mp3
>>
File: Rainbow kek.png (287 KB, 500x500)
287 KB
287 KB PNG
>>38772645
>Mares! NovelAI:
LOL the voice crack, holy shit
>>
>>38772645
>https://u.smutty.horse/mhwrjdhkpki.mp3
LMAO
>>
>>38772645
>https://u.smutty.horse/mhwrjdhkpki.mp3
Found this voice crack hilarious.
Made a short lazy animation of it.
Enjoy ^-^

NovelAI Dashie - Mare Cluck
https://u.smutty.horse/mhwrqbovarw.mp4
>>
Dall-E just wont draw a good Dyx
Other attempts:
https://u.smutty.horse/mhwshtqqjod.png
https://u.smutty.horse/mhwshubjddx.png
https://u.smutty.horse/mhwshuftlgz.png
https://u.smutty.horse/mhwshtmpeqx.png
https://u.smutty.horse/mhwshtorjft.png
https://u.smutty.horse/mhwshtuneoa.png
https://u.smutty.horse/mhwshuibbio.png
https://u.smutty.horse/mhwshukfsjw.png
https://u.smutty.horse/mhwshudmtdd.png
>>
>>38772645
Made some more Content with NovelAI TTS, if you hear some artifacts it's mostly from audacity when I lowered the tempo on some of these clips.

Blue Mares when you ask them what they want to get for dinner:
https://u.smutty.horse/mhwsuzqjolj.wav

Discord is Funny:
https://u.smutty.horse/mhwsuyrmwoi.wav

Fluttershy Is a Homo:
https://u.smutty.horse/mhwsvaeuknh.wav

lululu:
https://u.smutty.horse/mhwsuxkoqgp.wav

Pinkie Panic:
https://u.smutty.horse/mhwsuwubwei.mp3

Spikes Thoughtful Poem:
https://u.smutty.horse/mhwsuxczwcf.mp3

Squeaky Show:
https://u.smutty.horse/mhwsuxnyomt.wav

Elegant Well Mannered British Twilight
https://u.smutty.horse/mhwsuwrlcpx.mp3

I'm really impressed with Fluttershy Is a Homo, I made a TalkNet version of this awhile back but I think this sounded a little bit better.
>>
File: 1612090.png (3.61 MB, 2000x2297)
3.61 MB
3.61 MB PNG
HOLY FUCKING SHIT!!!
I just found Applejack (Seed:bevy), and oh my god, this is hooves down the BEST "Yeehaw!" I've ever heard from an AI.
She sounds strange when there's no ! marks though, she sounds best with only exclamations. Still incredible!

NovelAI Applejack - YEEEEEHAW!
https://u.smutty.horse/mhwsvmonojy.mp3
>"Yee haw! Boy howdy! I sure am glad I'm a mud horse!"
>>
>>38773319
The second part sounds pretty good, but does that "Yeehaw" actually sound like Applejack? It kinda sounds like a generic Southern accent to me, but I don't know.
>>
>>38773335
It sounds just like her to me. Just to double check, found a canon AJ yeehaw.

For comparison:
https://youtu.be/LZO3LSe1_bg
https://u.smutty.horse/mhwsvmonojy.mp3
>>
>>38772645
Rainbow Dash has less of that fuzzy sound quality in novelai than 15
>>
>>38773374
IIRC the next update is supposed to remove that fuzziness but we'll see
>>
>>38773317
>https://u.smutty.horse/mhwsuwubwei.mp3
>https://u.smutty.horse/mhwsuwrlcpx.mp3
ded
>>
>>38773359
I could be wrong, so please correct me if I am, but isn’t this a case of overfitting? The sample and the video linked sound way too similar that it must be learning it off of those clips. Does the yee haw work for other characters too?
>>
>>38773530
Yep, pretty much any character can attempt it, and it's given similar strong inflections.
I feel like they've made it like... A specific function for key words where it'll be said a certain way regardless of character.
It just sounds way better with a more country sounding voice, so there may be some bias there still.
So yeah, you still may be onto something about that

A bunch of "yeehaws" from many different ponies I found + the smokiest anon you've ever heard, who I'm calling "Tarlung Anon".
>https://u.smutty.horse/mhwtpsdgoex.mp3
Ft. Cozy, Pinkie/Tiara Fusion, Spitfire, Applebloom/Sweetie Fusion(?), Granny Smith, Applejack, Tarlung Anon.
>>
>>38773688
I suspected that was the case, I’m still a noob to machine learning and AI so I wasn’t sure if I got the terminology correct :p
So i guess that means that these models are specially good at replicating lines from the show but not as good at making new lines?
>>
>>38767902
Claiming 4-26.

Also, this is probably an obvious one but I think it would be funny to reference the beginning of APPLES.MOV somewhere.
>Applejack, you can't buck all those apples.
>Fuck you I can't buck all those apples!
>>
As NovelAI's TTS is definitely a powerful multispeaker model, it's able to blend or combine attributes of different voices together, as a sort of voice fusion. Which honestly seems great for any works that could include character shipping and child conception with it's own voice.

It's ability to combine voices like this so cleanly gives so many benefits for creative works:
>Offspring voices
>Hybrid character voices
>Documentary/interviewer voices.
>Limitless voices for "anon" inserts.
>Characters attempting to imitate another character
>OC/background characters with voices that suit canon
>Changes to voices (Older/younger version of certain characters?)
>Voices for ponies of certain professions: Announcer, news pony, wrestler, etc.

There is a ton of potential here.
Thanks again to >>38771566 for pointing it out to us.

>>38773701
>Not good at making new lines?
It seems to do great at generating new lines just fine. It just seems to have additional biases so that lines already said before sound better than usual compared to the rest? Or something like that anyways.
>>
>>38773729
I could be wrong, but the fact that they were adding the mane 6 one by one in >>38771566 seems to indicate that it's NOT a multi-speaker model, because you'd need to only train it once, right? Maybe they're doing something different that we haven't looked into before.
>>
PPP helped me out a lot so I'll just ask here:whats the song you like the most from these threads?
>>
>>38773717
Keep an eye on who claims that scene.
>>
>>38773878
I wish there were more retro of the 30s-60s kino songs made, specially with the Rarity voice.
>>
>>38773878
Tendrils
>>
>>38771566
Just from a few tests, it seems that its quality is often as good as 15.ai while only needing one generation. Half of the time, it also gets the inflection and tone right. However, its quality suffers with long inputs, it can mispronounce words and the voices sometimes slip into a different accent.
>>
>>38773761
it's multispeaker but the speakers aren't selectable
the idea is that people generate their own voices by putting in random seeds and/or mixing them, sorta like artbreeder
but turns out it didn't generalize hard enough so it's pretty easy to stumble upon something close to a training-set voice, which is what people are trying to do
>>
>>38773965
What?
>>
>>38773729
in principle this could actually be done with almost any multispeaker model, the more speakers the better
or at least i've had decent success with making androgynous voices in the espnet demo by averaging random f+m pairs of speaker embeddings
though i think a variational autoencoder might be useful? all i know is it makes latent spaces smooth
alternatively get a fuckton of different voices so it has no choice but to generalize; one idea i've had is to scrape voice actor demo reels which would be delightfully ironic
>>
>>38774155
Link to song?
>>
File: Cum Hay Temp Still.png (414 KB, 1920x1080)
414 KB
414 KB PNG
"Cum Hay" - Ft. Anon, Applebloom and Applejack.
https://u.smutty.horse/mhwzpesshwk.mp4

Voices done by NovelAI TTS and NaturalReaders TTS.
Considering adding SFX and bobbing characters, but kinda already like it as it is thus far.
But yeah, enjoy the skit ^-^
>>
>>38775203
'The Tendrils - AJ FS'
https://u.smutty.horse/mefwbwqkiwq.ogg
>>
File: 1403063204828.png (1.06 MB, 1400x1008)
1.06 MB
1.06 MB PNG
Great panel, I've gone through it twice now to find something I misheard and it's a really solid and engaging five hours even on a repeat, nice work. I do still like how the panels aren't scared to get into the weeds and I thought the Q&A time was adequate.

Earlier I asked about something I thought I heard: 'reductive sequence' generation but this was something I misheard. After listening through again these are the things which I am curious about:

Recursive generation
Auto-regressive model
Gradient descent
(Symbol/Similar) ? replacement

I want to hear more about these things, if anyone can talk on them I would greatly appreciate it.

And if SortAnon is around- panelists recommended I check here to find you.
If you're around can you please explain a bit more about 'octaves' and their relationship to your work?
If anyone else wants to talk about octaves, please join
>>
>>38768169
Got most of the voices done. I’ll be working on the video later. Is there a certain barrier of quality I should pass for my clip specifically? I’m scared that whatever I try to make up will be seen as not good enough.
>>
>>38775204
is it just me or does applebloom kind of bleed into non-AB a voice occasionally?
>>
>>38775453
it's not just you, a lot of voices tend to sound like a weird blend of other voices most of the time. it's why getting multi-speaker models is so tricky.
>>
File: take a seat.png (249 KB, 600x522)
249 KB
249 KB PNG
>>38775204
>"Cum Hay" - Ft. Anon, Applebloom
>>
>>38775247
>Is there a certain barrier of quality I should pass for my clip specifically?
Specifically? Not really. Just follow the rules and you should be good.
>>
>>38775233
Is there full lyrics anywhere? Seems like a great song
>>
>>38775453
Yeah, sometimes Sweetie Belle bleeds into it, which is why it took forever to get the right lines to be as Applebloom as possible.

I'm certain there's a pure Applebloom in there somewhere, but it's near impossible to find cause of infinite voice combinations, being a Multispeaker model, and the only current method of voice selection being via the seed. Still, I think it got close enough and has a lot more emotional capability and intensity than her 15.ai counterpart.

I suppose if I really wanted more accuracy, I could try inserting her Novel AI lines into TalkNet at least once for a final pass.
>>
>>38775453
Ironic considering AB's VA was SB's singing voice for a while.
>>
I saw that some members of the koboldai team finetuned Gpt 13b. Does anyone here have access to a powerful enough machine to make a mlp 13b model?
>>
>>38777144
on the koboldai discord the guy responsible for it said one a6000 can do 0.26 samples per second (with optimizations), and his ~1gb dataset seems to have ~129k such samples https://www.buymeacoffee.com/mrseeker/finetuning-picard-v2-fairseq
so that's ~138 hours of a6000 time, ~$75 on vast.ai
which sounds so low that i think i screwed my math up
>>
>>38777555
and that's for on-demand, interruptible is half the price
the reason i'm questioning this is because the linked post quotes $45 for a *2.7b* model

also realized that the post talks about a 6.7b model but i'm only getting dataset size from it so it doesn't matter
>>
>>38776333

The Tendrils
[Verse 1]
A scourge born from the forest
Under fog and haze
Crept towards the town as
They rested from a busy day

Sun and moon lit up the sky
Our minds were in a fray
Something just wasn’t right
Something met us on our way

[Chorus]
The tendrils sprouted from the ground
Choking all their prey
Whatever ponies they had found
Are trapped in viny cage

The tendrils on a savage path
For all the trees we pruned
They plundered all we had
The tendrils’ wrath we knew

[Verse 2]
The sky was streaked with lightning
That pegasi couldn’t tame
The toughness of these vine things
Put shame to the Apple name

Neither could the unicorns
With all their magic spells
Halt the plunder with their horns
Their sorcery had failed

[Chorus]
The tendrils sprouted from the ground
With nothing we could do
Only Twilight could save us now
Without her we were through

The tendrils on a savage path
For all the trees we pruned
They plundered all we had
The tendrils’ wrath we knew

[Verse 3]
Our saving grace, or so it seemed
A cloud-white liquid brew
The princess took a little drink
To seek a hidden truth

To tell the truth, I’m not quite sure
What Twilight had seen
But we had finally found the cure
To save a Tree of Harmony

[Chorus]
The tendrils sprouted from the ground
With twists and turns and bends
They did their deed and withered out
And met a timely end

The tendrils on a savage path
For all the trees we pruned
They plundered all we had
The tendrils’ wrath we knew
>>
>>38775233
do you think AJ's voice would have a stronger southern drawl if the reference audio had one?
>>
>>38770759
Hmm, the original video is 35s, but the joke theme Im going for is requiring me to extend it (for at least 10 extra seconds, maybe possibly 5 more)
>>
>>38771568
I brute force tried a lot of seeds. Here's what I came up with:
Apple Bloom 0x533cf
Applejack 0x7336
Princess Celestia 0xa7a4
Discord 0x6527
Fluttershy 0x62a66
Granny Smith 0x1159d7
Pinkie Pie 0xb2a6d
Rainbow Dash 0x13c39
Rarity 0xb664e
Spike 0x1041e4
Sweetie Belle 0x10b00
Trixie 0x102816
Twilight Sparkle 0xd579b
>>
>>38767902
>>38770759
BGM, could I bother you with sharing the HD version of that image from last redub op, the pink one with "look Lauren Faust, we are also doing ponies!"?
>>
>>38777972
Samples and added Scootaloo:
Apple Bloom 0x533cf
>Well, one thing is for sure. The Cutie Mark Crusaders will never be the same. We'll be better!
https://u.smutty.horse/mhxjsgmnxhu.wav

Applejack 0x7336
>Yee haw! Finally I can buck like a five-bit snake herder in an Appleloosa ranch house again!
https://u.smutty.horse/mhxjsljomna.wav

Princess Celestia 0xa7a4
>Hm, a very valuable lesson to have learned.
https://u.smutty.horse/mhxjstsntgp.wav

Discord 0x6527
>I'm not playing fair? Perhaps we haven't met. I'm Discord, spirit of chaos and disharmony. Hello?
https://u.smutty.horse/mhxjsuipgrz.wav

Fluttershy 0x62a66
>Um, excuse me, sir. I mean no offense, but your rhythm is just a teeny tiny bit off.
https://u.smutty.horse/mhxjstwshdv.wav

Granny Smith 0x1159d7
>I love me some nachos. I'mma set my dentures loose on these vittles until the crumbs cry uncle!
https://u.smutty.horse/mhxjsubseih.wav

Pinkie Pie 0xb2a6d
>Well, that just takes the cake. Wait, no! It can't take the cake 'cause there is no cake! Or muffins! Or griffon scones!
https://u.smutty.horse/mhxjsuoxaud.wav

Rainbow Dash 0x13c39
>It needs to be about twenty-percent cooler.
https://u.smutty.horse/mhxjstplqoz.wav

Rarity 0xb664e
>I used to be somepony. I used to be respected! I made dresses. Beautiful, beautiful dresses. But now, everypony is laughing at me. I'm nothing but a laughing stock!
https://u.smutty.horse/mhxjsugnehe.wav

Scootaloo 0xad595
>Winners or not, we still have the chance to be awesome!
https://u.smutty.horse/mhxjstuoofp.wav

Spike 0x1041e4
>This is hilarious! Look at all of you! We got: Hairity, Rainbow Crash, Spitty Pie, Apple Teeny, Flutterguy, and... uh... I got nothin'.... Twilight Sparkle. I mean seriously, I can't even work with that...!
https://u.smutty.horse/mhxjsumwwfw.wav

Sweetie Belle 0x10b00
>I don't think Rarity would let me travel to the far reaches of Equestria for cutie mark problems.
https://u.smutty.horse/mhxjstzaypy.wav

Trixie 0x102816
>You may have vanquished an ursa 'minor,' but you will never have the amazing, show-stopping ability of the Great and Powerful Trixie!
https://u.smutty.horse/mhxjsuktrwx.wav

Twilight Sparkle 0xd579b
>I'm doing scientific research. I'm observing Pinkie Pie, scientific name: Pinkius Piecus, in its natural habitat.
https://u.smutty.horse/mhxjsuelsbo.wav
>>
File: 1630026858506.png (156 KB, 896x504)
156 KB
156 KB PNG
>>38778137
>>
>>38777907
It's possible, maybe listening to country singers and the way they perform could highlight
>>
>>38767902
Claiming 4-7
>>
I want to spend some time roadmapping the next few months of stuff on my side. Some candidates...

IPFS stuff:
- Have my servers use ipv6 rather than ipv4 for faster downloads.
- Set up an IPFS gateway on my servers so it's easier for people to navigate pony files on IPFS.
- Improve my downloader so it's easier for people to download things from IPFS.
- Learn to use IPFS's UnixFS and organize the IPFS dataset better.
- Clean up, document, and publish scripts for managing IPFS datasets.

Animation stuff:
- Upload the XFL files.
- Make it easier to find assets in the XFL files.
- Explain how to pose and animate assets in the XFL files.

Animation AI:
- Start exploring animation AI, maybe for asset generation, keyframe generation, tween generation, or frame composition.

Other datasets:
- Update & publish my image metadata file for linking images across datasets.
- Automate getting dataset updates for derpibooru & desuarchive.
- Scraping /mlp/ directly without depending on desuarchive for future posts.
- Scrape more boorus.

Dataset tooling:
- Create a local version of the data-subset creation tool.
- Create image & post data-subset creation tools.
- Make my datasets compatible with WebDataset.
- Make my datasets more accessible to cheap compute (vast.ai or other).

https://poll.horse/hzkr4YYK
I probably won't follow the poll results directly since some things are prerequisites for others (e.g., a lot of data tooling depends on IPFS stuff), and since some things will take way more effort than others. But I want to know what everyone wants to prioritize.

>>38763817
Can we meet tomorrow over the cytube to break these down into timelines and milestones?
>>
>>38775204
I was waiting for AJ to beat AB. Why you do this to me, Anon?
>>
>>38775204
>implying AB wouldn't compete with AJ to see which one makes the bigger mess
>>
>>38775244
Thanks!
- Recursive generation: This was probably in reference to animation assets (symbols). Each asset is represented as a 2D grid of layers and frames. Each cell of this 2D grid is a composition of other asset frames. The fact is each asset is composed of other assets is what makes this recursive. If we want AI that generates assets in a similar way, that AI would have to generate assets recursively.
- Auto-regressive model: This refers to a type of neural network that generates sequences. The idea is that the neural network maintains some internal state while generating the sequence, and the sequence items it generates are tied to its internal state. There's a good picture of it here: https://ml.berkeley.edu/blog/posts/AR_intro/
- Gradient descent: This is the standard algorithm used to improve neural network parameters. The name comes from an analogy with how the loss function (= how badly the network performs) is represented. People visualize the loss function as a landscape with a bunch of hills and valleys. The x, y position (equivalent of GPS coordinates) represent the neural network parameters, and the height of the landscape at that point represents the value of the loss function. The goal is run downhill, which, according to the loss function, makes the network less bad. Running downhill in the analogy represents changing the neural network parameters iteratively. The "direction" of the change in parameters can be calculated using a (calculus) derivative.
- Symbol/similar replacement: I'm not sure what this is. Where is it mentioned in the talk?
>>
Is TalkNet down/gone? I can't access it anymore.
>>
>>38775244
>If you're around can you please explain a bit more about 'octaves' and their relationship to your work?
I'm not SortAnon, but I can explain a bit about octaves.
Music is generally written to a certain "key" which corresponds to a sequence of notes that sound good together. The key is described by (1) a letter to describe the root note, one from A through G, (2) an optional modifier for the root note, either flat or sharp, and (2) a scale modifier, usually either major or minor. For example "A sharp minor", or "B flat major", or "C major". The root note letter and its modifier describe some starting pitch for the scale. The scale modifier tells you which notes typically fit well together. A typical scale would look like this:
- C, D, E, F, G, A, B, C
- A, B, C sharp, D, E, F sharp, G sharp, A.

If you want to play notes higher or lower, you continue in a cycle.
- C, D, E, F, G, A, B, C, D, E, F, G, A, B, C, ...
An octave is the difference between C and the next C in this cycle. (Or between A and the next A in this cycle.)

A letter + accidental (accidental = letter modifier) together pick out some subset of notes.
HOWEVER, even though these three things are enough to describe the key of some music, they are NOT enough to describe a specific set of notes. That's because the scale cycles, so a single letter+accidental corresponds to one note for every cycle. This is important because different instruments can't all play exactly the same notes (some play too high, and some play too low), and what you really want is a way to describe music so that (1) one instrument can substitute for another if needed while still sounding like basically the same song, and (2) different instruments can play together without sounding terrible. The only reason it's possible to do this is because when you play a musical part off by an octave (too high or too low by exactly 1 cycle), it often still sounds good.
That's what SortAnon is using for TalkNet. He's getting the AI to play the correct letters+accidentals, but he's letting the AI choose to go up or down a cycle based on what works best for it.
>>
>>38778603
Delivering 4-7
https://u.smutty.horse/mhxluggdsdj.mp4
>>
>>38779046
It's working for me.
>>
>>38779112
The scene after that, whoever does it, could make a reference to the "I came paper" image. If the joke works.
>>
>>38767902
>>38768169
I'm almost done, but I've had a problem with one of Big Mac's lines. This line specifcally:
>But still only one pony, and one pony plus hundreds o' apple trees just doesn't add up to...
It's not exactly what he says, but how he says it. He has a slight pause after saying "pony", and slows down when he says "hundreds". How exactly should I deal with this?
>>
>>38779236
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
xarray-einstats 0.2.2 requires numpy>=1.21, but you have numpy 1.19.5 which is incompatible.
datascience 0.10.6 requires folium==0.2.1, but you have folium 0.8.3 which is incompatible.
albumentations 0.1.12 requires imgaug<0.2.7,>=0.2.5, but you have imgaug 0.2.9 which is incompatible.

This comes up. Seems like the only error, but no idea how to fix. Soon as it loads, I get a 403 error. Does the collab page no longer work and TalkNet is only offline now?
>>
>>38779282
I get that error too, but it still runs fine. If you see a "Successfully installed numpy..." after that, then try just running the next cell.
>>
>>38779272
Splicing audio? What are you asking?
>>
>>38779420
Generating the audio through 15.ai
>>
>>38779112
It's great Pinkie Pie is in fine mental health and there's no reason to suspect she's a danger to herself or anyone around her.
>>
>>38779424
You can edit the audio afterwards right?
>>
>>38779424
Using different punctuation marks can help break up sentences, or splice different takes together. You can also stretch words out in your daw with whatever time stretch feature yours has
>>
I've finished editing the upgraded version of the PPP panel with all the raw voice and video recordings, will render overnight and post tomorrow, all going well.

>>38779000
Happy to discuss live tomorrow, should be free anytime after 1900 BST.

Here's my current to-do list, in rough order of priority:
>Finish and upload PPP panel recording
>Add recent dataset upgrades, GPT-PNY and NovelAI to the quickstart guide and update the active tasklist therein
>Sort out the current accessibility issues with (and maybe cull/re-write parts of) the main doc
>Ensure all tools, datasets and resources are up to date and easily accessible
>Make some sort of easily sharable pic detailing all the PPP's datasets and current available AI tools, then spread it around the board
>Maybe some dataset stuff, depending on what we end up deciding tomorrow
>At least one contribution to the ep.4 redub
>Make a new large-scale AI thing
>>
>>38779463
>>38779483
I don't know how to make it sound good. Please help.
https://u.smutty.horse/mhxnqmxneir.flac
>>
>>38779676
1900 BST works for me.
>Make some sort of easily sharable pic detailing all the PPP's datasets and current available AI tools, then spread it around the board
That's a good idea. If you don't mind waiting a couple of months for this, I'd like to get my IPFS stuff and dataset tooling stuff sorted out before you work on this. That way, we can have everything on the data side be clean and easy-to-use.
>>
>>38779790
That's 15.ai? sounds different from when I use it. Either way generally the easiest way to change the pacing of a line delivery is to add commas and try different emotional contextualizers on the input when you're generating.
>>
>>38779790
I don't know what you did but it's got me laughing pretty hard so thank you for that. You could try rendering one sentence at a time and sticking each together again, and messing with the emotional modifiers.
But after testing a few myself I've come to the conclusion that Big Mac just sounds like garbage. Maybe embrace it, it's pretty funny as is
>>
Is it normal for the master file mega links to take a long time to load? I'm trying to get a specific sound effect but it's taking too long and I'm not exactly sure which link it is.
>>
>>38780165
Mega doesn't handle large folders well. It's better to download the master files (ideally with their dl manager, less of a chance something breaks) and search locally. Searching by name is another thing that's trivial locally but not possible on mega.
>>
Can multispeaker mix CMC to make an anon filly voice if possible?
>>
>>38779960
It's 15ai, but he slows down the clip after Mac says "but still only one pony,".
>>
>>38767902
Delivering 4-26, with some animation by FluttershyFan.
https://u.smutty.horse/mhxpaxxjjpr.mp4

>>38779790
Also are you going for a 1-1 copy of the original dialogue? Just so you know, you can change the lines to make them say whatever you want, that's half the fun.
>>
Claiming 4-12
>>
what do you guys think of commercial TTS getting closer and closer to 15.ai's standard?
>>
>>38780685
post examples
>>
>>38780692
see novelai pony tts
>>
>>38780697
It's not there yet, see above
>>
>>38780685
wait until you see what the governments have
>>
Here with a nightly check in on Dall-E
Took a bit of coaxing but it finally made me a profile picture
Anyway, while getting it to make deliberate ponies is a pain in the ass, it seems to be fully capable of making FiM style OCs, at least in oil painting style
https://u.smutty.horse/mhxqhpqakwa.png
https://u.smutty.horse/mhxqhpwiaut.png
https://u.smutty.horse/mhxqhpyhwas.png
https://u.smutty.horse/mhxqhpuhsll.png
https://u.smutty.horse/mhxqhpsfflh.png
https://u.smutty.horse/mhxqhqcojto.png
https://u.smutty.horse/mhxqhqnayzt.png
https://u.smutty.horse/mhxqhqaexrn.png
https://u.smutty.horse/mhxqhqfmhbn.png
https://u.smutty.horse/mhxqhqhwmpn.png

This last one is particularly interesting since it actually looks like RD, Chryssy, Pinkie, and AJ
https://u.smutty.horse/mhxqhqkkunt.png
>>
>>38780708
not yet, of course. But its getting closer and closer. What might be happening 6 months from now?
>>
>>38780720
15.ai hasn't updated in almost a year now, so it's actually the standard set back in 2021
>>
>>38780316
KEK
>>
>>38780724
I'm looking forward to when 15 releases his update he's been working on for months
>>
>>38780715
I'm concerned about the obvious eqg in the dataset, it will never even be a toggleable switch in their products
>>
>>38780715
>https://u.smutty.horse/mhxqhqfmhbn.png
mmmmmh that pony ass
>>
>>38780736
yeah, those were the cream of the crop
https://u.smutty.horse/mhxqktoldnk.png
is one of the worst cases of barbie syndrome in the current generation round
>>
>>38780715
>Griffin
>Draws hippogriff instead
Future is not here.
Anyway, how did you use it? I thought Dall-E 1 wasn't available to public.
>>
>>38780756
waitlist. I'm another anon, but I got an email as an artist to join an onboarding session about a day from now.
>>
>>38780715
In the prompt engineering doc I can see why its not that coherent when there's a lot of ponies on screen:
>4: More than one character/objects in a scene
>4.1: DALL·E 2 has no problem with creating more than one person/character of interest in a scene, but falls short in separating certain characteristics of each from each other, choosing to mish-mash whoever you mention into a totally different set of personas, or “create” new ones.
The best thing to do is focus on the background, then use inpainting for the rest of the ponies in the positions where you want them to be.
>>
>>38780756
>>Griffin
>>Draws hippogriff instead
that's a funny bit
>AI doesnt seem to know what a hippogryph is and kept throwing hippos at me
>I wanted a hippogryph anyway since the OC is a hippogryph and not a griffin
so Dall-E failed the task successfully, likely due to it merging griffin and "my little pony" into a single entity
Here's another hippogryph it made when I told it to make a griffin
https://u.smutty.horse/mhxqnhwlxht.png

One quirk I found is that it also sucks at making just griffons, let along, show accurate ones, since it has a heavy bias towards the griffon vulture instead of the mythological griffin, and even then likes to use the eagles and griffons from crests and heraldry instead of actual art and models from tabletop and video games
>>
>>38780801
Maybe it got confused because FiM doesn't use the word "Griffin", it uses "Griffon".
>>
>>38762176
1 / 2
I managed to download one of the render traces. Is there more information on what the render trace format is? I thought there might be something in the /mlp/ con panel although I didn't notice anything on a quick skim.
https://www.youtube.com/watch?v=lQsRoRxFsfw
It seems like you have code to render xfl.
https://github.com/synthbot-anon/synthanim
Can render traces be rendered directly or converted to something that can be rendered? An AI could spit out this kind of data although if I can't see the result it would be hard to work with.

This is my current understanding of the render trace format feel free to correct me if I'm wrong. In a render trace directory there's some files frames.json, labels.json, logs.txt, and shapes.json. It seems like frames.json maps numeric ids to other numeric ids and transform matrices. It seems like labels.json maps numeric ids to layers containing shapes and to a specific frame. It seems like shapes.json has all the layers and the shapes in the layers with rendering information like line size and whatnot. So putting it all together it seems like I can access all the shapes in a specific frame. Then I could apply the transform matrix of a layer at a specific frame to move the shapes into where they're supposed to be at a specific frame. Is there something important that I'm missing?

Could you give a rough estimate on how close the data render is to the show? For now it seems like the data could be worked with although in the future it might be improved with more cleaning. Another point is generting meaningful labels for the animation dataset. For now it looks like it's just in one big data pile with some names like octavia_leg and whatnot. There's research on generating a large set of good labels from a small set of about 30 manually labeled images.
https://nv-tlabs.github.io/datasetGAN/
>>
>>38780827
2 / 2
Also this is what I did to download from ipfs if anyone else is trying. The http url is slow if it works at all. The browser plugin is slow if it works at all. I ended up doing
ipfs init
ipfs swarm peering add /ip4/135.181.60.95/tcp/4001/p2p/12D3KooWLMr455Va1fH5XxX8EJXHJFQaSgMyaU2YzSryzV8ujBaX
ipfs swarm peering add /ip4/176.9.11.137/tcp/4002/p2p/12D3KooWMHxW4x1Dp3rjXf3UxKpH9u7XTgBfu5gzCCdyMWjHkBCg
ipfs swarm peering add /ip4/176.9.11.137/tcp/4003/p2p/12D3KooWNJCmwFGFNZGzeCeWxasrckMmJxfLEqrG6AxzSPJ4NSWd
and then used the webui url that running
ipfs daemon
printed out. Alternatively I could use ipfs-desktop which handles the daemon. In the webui I added the directory using the id
QmdMjH7EsHdd4gGgCUnssDWndf54rVXQANvaSZFnhp5Tnw
from
ipfs://QmdMjH7EsHdd4gGgCUnssDWndf54rVXQANvaSZFnhp5Tnw
Accessing the directory directly was slow. It was a lot faster to right click the folder and choose inspect which I presume lets me access the directory without loading all the metadata. Once I found the id for one of the render trace folders I did
ipfs get QmNMwAiGwCVxXoWJGE8SH6FQYck2ZRrdeEo22x6EXV3fuq
which downloaded the folder to the current local directory. Browsing the files is a bit annoying. Is there a list of render trace ids that I could run ipfs get on?
>>
>>38780767
>waitlist
Oh, so it's real? Usually it's a nice PR-friendly way to say "never". Unless you're famous or something.
>>
>>38767902
Delivering 4-18
https://u.smutty.horse/mhxqreqvcgl.mp4
>>
>>38780833
kek
>>
>>38780831
I suppose I could download the 150 gig directory or some of the 10 gig subdirectories. Since the ipfs ids are immutable a change in a dataset file would correspond to a different ipfs id so I suppose a list of ipfs ids might not be a good idea.
>>
Gained trail access for MidJourney a little while ago after signing up a couple days ago, namely thanks to Jacksepticeye mentioning it.
Damn it's so chaotic to view your own images, because it's bot is embedded in their discord and multiple users share a channel.
That being said, it's capable of making some decent results, and amazing ones if they're more artistic based.
It has a hard times with characters, but an easy time with environments.

>Rainbow Dash standing on a cloud (image related)
Full result - https://u.smutty.horse/mhxqvdhbwko.png

>My Little Pony: Friendship Is Magic
Low% batch result - https://u.smutty.horse/mhxquqhoukz.webp
Full batch result - https://u.smutty.horse/mhxquqrivks.png

>My Little Pony reimagined by Akira Toriyama
Batch result - https://u.smutty.horse/mhxqvmbphnf.png
Cursed Diamond Tiara - https://u.smutty.horse/mhxqvnrahkr.png
Cursed Pinkie Pie - https://u.smutty.horse/mhxqvsokwgn.png

>An expressive oil painting of the Australian outback on fire in rainbow colors
Batch result - https://u.smutty.horse/mhxqwbnekla.webp
Favorite upscaled to max - https://u.smutty.horse/mhxqweitpal.png
>>
>>38780867
I'll take those prompts and run each of them through Dall-E tomorrow afternoon/evening
>>
>Borg assimilated My Little Pony
Batch result - https://u.smutty.horse/mhxqwvtovdr.webp
Upscaled #2 (image related) - https://u.smutty.horse/mhxqwwitokl.png
Upscaled #3 - https://u.smutty.horse/mhxqwwcziqx.png

>Pinkie Pie trapped inside a crystal ball
Low% batch result - https://u.smutty.horse/mhxqxezwsfo.webp
Full batch result - https://u.smutty.horse/mhxqxofocez.webp

>A world where people are quadrupedal horses, as viewed from on a subway
Full batch result - https://u.smutty.horse/mhxqyewxioq.webp

>Lyra and Bon Bon sitting on a park bench
Low% batch result - https://u.smutty.horse/mhxqymrfvuw.webp
Full batch result - https://u.smutty.horse/mhxqyrbnqcf.png

And finally, another user in the same channel generated this and it's hilarious.
>Pink space poodle riding a retro rocket
Full result upscaled - https://u.smutty.horse/mhxqzfxazeq.png
>>
>>38780895
[Missing image related]
>>
File: 217901.gif (3.58 MB, 576x324)
3.58 MB
3.58 MB GIF
>>38780895
>Full batch result
>>
And remember, only the Gothiest of ponies coat themselves in raw crude oil.

>Goth My Little Pony
Found from another user (Max upscaled)
>>
>>38780867
i believe with midjourney you can generate invites yourself to bring other people in?
>>
File: MidJourneyCosts.png (61 KB, 1090x742)
61 KB
61 KB PNG
>>38780944
>you can generate invites yourself to bring other people in
Yes, but the invites are only to give others free trials when you pay for a plan. The free trail image amount is about 27 currently, upscaling and variation counts as another image.

"Friend passes allow you to invite friends to the server and give them a limited number of free images they can make. Everyone starts with three friend passes, then we give out extra every month."

I'm definitely gonna get at least a basic plan for this, so I'll get a couple.
I may share at least 1 with one of my personal friends who may be interested in it. Otherwise I may share the rest here either immediately or occasionally for interested anons.
>>
>>38775233
very nice, I like the groove
>>38780715
pretty. well, they're mostly pretty until you look closely
>>
>>38780316
I love this, many great callbacks to older episodes too.

>>38780715
Woah, a lot of these look amazing! A nice PFP too, I really like their mane, and Hawaiian shirt..

My favorite of the others is still "https://u.smutty.horse/mhxqhpqakwa.png" but I also like "https://u.smutty.horse/mhxqhpwiaut.png", "https://u.smutty.horse/mhxqhqfmhbn.png", and "https://u.smutty.horse/mhxqhqhwmpn.png"
>>
>>38779790
Do you want to try recording the line yourself, then using TalkNet to transform your voice? I still wish more people used TalkNet for acting instead of singing. If you're a good enough actor, you can control emotion and delivery in a way you can't get from 15.

The problem with using Big Macintosh is that all of his current models sound too high-pitched because there's too much "Orchard Blossom" in his dataset.
>>
>>38780827
https://ponepaste.org/7569
We can compile render traces directly to SVG. So if we get AI that can produce render traces, we're basically done. My code needs to be cleaned up a bit so it's more clear how to work with everything. There are command-line options for converting XFL to render traces, then for converting render traces to SVG. I included them in the paste.
>This is my current understanding of the render trace format feel free to correct me if I'm wrong.
Check the "Understanding timelines and layers" section of the paste. frames.json describes a DAG structure for all frames in a scene. shapes.json describes the leaf nodes of that DAG. labels.json tells you how to go from timeline tables to frames.json identifiers.
>Could you give a rough estimate on how close the data render is to the show?
- The gradient fills are incomplete. This is very noticeable when it occurs because it results in a shape not getting filled. It affects a few character assets. You can get an estimate for it by scanning through these animations: https://drive.google.com/file/d/1YKMjr6SmZcMMNedT2L5edvsiWCaN9Kar/view?usp=sharing. The graininess is caused by ffmpeg, not by any issues with the render traces. In places where a character looks deformed, that's almost certainly a problem with the original data, not with the render traces. These are all assets in the "noisy" pile of characters.
- The way some lines connect is blurry. This is barely noticeable.

In both cases, the render traces are perfectly accurate as far as I know. If there are any bugs in the render traces, it's a bug I don't know about. All of the problems right now are for compiling render traces to other image formats, so any AI working with and generating render traces directly should be unaffected.
Future efforts on data cleaning would be focused on labeling show assets as "noisy" or "clean". Clipper made a first pass of this based on a still image of around 10,000 assets. In the (potentially near) future, he might want to do it for the rest of the show assets (maybe including backgrounds). We'll also want to take the ones he's labeled as "clean" and fully render them so we can make sure that the full asset is clean.

>>38780831
>>38780856
>The browser plugin is slow if it works at all.
Once you run ipfs daemon and add the swarm peers, the browser plugin should run a lot faster. It uses your local ipfs daemon if available.
>Is there a list of render trace ids that I could run ipfs get on?
ipfs get /ipfs/QmR3sVyb3ZWmiESWXfsTndktJ3ZE77rokPifMBJwrjNPEE
That will download all of them. It's about 131 GB. Here's one of the 10 GB chunks:
ipfs get /ipfs/QmbPgFW6a71kjfySi4u45TabCYqGVWtJ3JqFncaVLkaqRL

This data is supposed to be immutable, so it's not so bad to use content ids. At some point, I'll set up some script for dealing with IPNS for mutable data.
>>
NEW THREAD
>>38781041
>>
File: TrixQT.png (349 KB, 640x720)
349 KB
349 KB PNG



Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.