General dedicated to creative use of free and open source text-to-image models.Previous /ldg/ bread : >>101128616Agreeable Edition>Beginner UIFooocus: https://github.com/lllyasviel/fooocusMetastable: https://metastable.studioEasyDiffusion: https://easydiffusion.github.io>Advanced UIAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiStableSwarmUI: https://github.com/Stability-AI/StableSwarmUIInvokeAI: https://github.com/invoke-ai/InvokeAIComfyUI: https://github.com/comfyanonymous/ComfyUI>Auto1111 forksSD.Next: https://github.com/vladmandic/automaticForge: https://github.com/lllyasviel/stable-diffusion-webui-forgeAnapnoe UX: https://github.com/anapnoe/stable-diffusion-webui-ux>Use a VAE if your images look washed outhttps://rentry.org/sdvae>Models, LoRAs & traininghttps://civitai.comhttps://huggingface.cohttps://github.com/Nerogar/OneTrainerhttps://github.com/derrian-distro/LoRA_Easy_Training_Scripts>Pixart Sigma & Hunyuan DIThttps://huggingface.co/spaces/PixArt-alpha/PixArt-Sigmahttps://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiTComfy Nodes: https://github.com/city96/ComfyUI_ExtraModels*SD.Next also works with PixArt-Sigma>Animationhttps://rentry.org/AnimAnonhttps://rentry.org/AnimAnon-AnimDiffhttps://rentry.org/AnimAnon-Deforum>Index of guides and other toolshttps://rentry.org/sdg-linkhttps://rentry.org/rentrysd>View and submit GPU performance datahttps://vladmandic.github.io/sd-extension-system-info/pages/benchmark.htmlhttps://docs.getgrist.com/3mjouqRSdkBY/sdperformance>Try online without registrationtxt2img: https://www.mage.spaceimg2img: https://huggingface.co/spaces/huggingface/diffuse-the-rest>Share image prompt infohttps://rentry.org/hdgcbhttps://catbox.moe>Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
Blessed thread of frenship
>>101146316>he made the collage
>>101146316nice collage
>>101145954>model to be trained ENTIRELY on AI SLOPHAHAHAHAHA
>>101146350>he didnt make the collage (this time)
>>101146389it probably works tho
>>101146389It's over. >>101146413>works
>>101146389>ethics>inbreeding AI pretrainingit's over before it even startedMeanwhile at gigachads OpenDiffusionhttps://new.reddit.com/r/Open_Diffusion/comments/1dn7h53/open_diffusion_mission_statement_10/>The goal of Open Diffusion is to create Open Source resources and models for all generative AI creators to freely use. Unrestricted, uncensored models built by the community with the single purpose of being as good as they can be.
>>101146389>he views it as a challengeI mean, all things aside, that's kinda based. Won't be much if he's proven wrong though.
>>101146316W000 finally made it in to a collage!
>>101146487
>>101146478he will waste time and money for nothing, has any SDXL finetune find success by only being trained with AI pictures?
>>101146456AI has random mutations if the selection is reasonably solid there is no inbreeding anyhow.I don't see the issue.
>>101146498>he will waste time and money for nothingvery likely, but hopefully we're wrong
>>101146505there was a paper showing that the more you train a model with AI pictures, the more artifacts you're getting, it's not a good idea at all
>>101146505>AI has random mutations. If the selection is reasonably solid there is no inbreeding anyhow.That's also a fairly good point. Question is, how good is their selection going to be. Last time I looked into a laion dataset it looked worse than /b/.
I simply do not wish for my outputs to look like AI therefore I do not train on AI
>>101146498If it's SDXL that may be a mistake.>>101146527There's no way. They probably did a variant without selecting images between trainings or just for an extremely limited scenario.
>>101146389This is fucking retarded. Okay fine, you don't want to scrape the internet. How about, then, have people upload pictures they have taken with their cameras and require metadata with those images. Add on top of this proper public domain media and cc0 images and you will have tens of millions of images.
>>101146529And Civitai is probably a better dataset for 1girl, anime and comparable, though perhaps not for all things that were done in human artwork so far.But it just needs to do a model anyhow, not the model to end all models?
>no new poses>no new locations>noise and watermark data from dalle gens>no new actions or interactions>no new anglessabotaged from the get go
>>101146603what would be the point of training something if it is not better than the thing preceding it
>>101146599Events and trademarks and people and object designs inside the photos and so much more can be intellectual property of someone else than the photographer thanks to copyright lobbying making endless amounts of things copyrightable/trademarkable/patentable between nations.
>>101146651this, if you train a model with only AI pictures from an another model, the best scenario would be the new model would be 90% as accurate as the the other AI model used to make the dataset, that's dumb as fuck, AI should aim to learn real pictures, not AI not so accurate pictures
>>101146651They're retarded because they don't realize the opponents to AI are doing it in bad faith because at the end of the day: IT IS ABOUT MONEY.
>>101146651Who says it won't be better? It can be better by better being able to do some combinations of concepts or by prompt adherence or anything else.
>>101146389they haven't learned anything, when SAI decided to add an op-out for the artists, they weren't loved for that, the artists don't care, they want AI gone, period, you'll never appease those motherfuckers so fuck them
>>101146663This is nonsense. Sure, you maybe can't take a picture of the comic book you own, but you can absolutely take a picture of your sonic plushie. And you can absolutely walk down Times Square in New York and record it with a high quality 4K camera, produce 100k great images from it, and be completely free legally.
>>101146683>Who says it won't be better?it definitely won't be better aesthetically, that's the problem, if you give to an AI model only AI pictures, you're basically telling it that "this is the most beautiful pictures you could aim"
>>101146527>the more you train a model with AI pictures, the more artifacts you're gettingThe issue is twofold. If the selection method is highly automated (which it likely is) then I'm not optimistic about it. A good dataset taking from synthetic materials would require manual labour of keen eyes, both for visual artifacts, inconsistencies, as well as aesthetics.
>>101146672It can actually be more accurate than the last imagegen AI by technology improvements, further tagging or something else. It can also be more aesthetic or more diverse on average by curation of many partly randomized variants encouraging it to be.Human made artwork operates on less.
>>101146727You seem to be blind, not a good outlook. Blind and retarded.
>>101146707No, Sonic is an intellectual property of Sega. You need to take and use the photo in the right country to make this possibly viable for a specifc use.
>>101146727>Human made artwork operates on less.a machine learning architecture isn't a human brain, all it does is to try to copy the pictures you've given to it, and it will never be 100% accurate, so it will make pictures less accurate than the AI pictures used for the trainingBesides, the human brain has been trained on a shit ton of data, let's suppose your eyes can see at 24fps, it means that you're being trained on new pictures 24 times per second for years and years, no AI would have been trained as good as us
>>101146727Human made artwork have infinite more value because the mere act of someone creating it gives it value because that person spent time on it. AI art is shit by default especially the shit you see spammed on Civitai. I can't imagine making a model from the outputs of Civitai.
>>101146718Only if there is no selection process in between. It's indeed not going to work without the selection. Same as for biology, if you just have mutations and more mutations with no selection (though no selection is impossible in physical reality) the results would definitely be stupid.
>>101146755If you can monetize it on Youtube it's okay. I can record a video of my Sonic plushie and upload it to Youtube and make money. Your idea of what is safe and not safe is infantile but I'm sure corpos would love your asinine take.
>>101146651most of ai art is shit, but that doesn't mean there's no good material it good learn fromai by it's very nature is transformative, meaning it's capable of creating new, good works, and even traditional artists have their flawsthere's a reason both AI and humans hide their hands
>>101146824Go ahead and train your shitty AI-image model. lmao
>>101146791for drawing it's debatable, for real life pictures that's a retarded thing to do, no AI pictures will be 100% accurate on anatomy, perspective, and lightning, so you're basically training an AI on non accurate outputs, guess what results you're gonna get? probably shit
>>101146841I'm just making a point, no need to project.
>>101146845There is no point, AI-generated images are ass and only copious amounts of cherry picking and inpainting might get you at a point still worse than an average competent artist's work.
>>101146856I'm not disagreeing with you..
>>101146824i think pixart trained on mainly mj and it turned out ok and animagine also clearly trained on nijibut i cant see how its going to ever be better than what we have if those retards train ONLY on that
>>101146813Good luck uploading your commercial video featuring 100k copyrighted / trademarked / design patented (USA) items at a intellectual property protected event location.
>>101146909Hell, there's been plenty loras trained on outpust of previous models due to some of them having a nice and distinctive style, and they turned out well too If it works on a smaller scale, it COULD work on a bigger one, if with major adjustments. At the end of the day it's about quality and plurality of source material.
>>101146875The problem with using AI-based images is there's the temptation of laziness, to use millions of slop AI images with little to no filtering. Whenever anyone talks about using AI they're always thinking about that aspect. "Boy it's annoying to get millions of images, why don't I just generate them and I get the prompts for free..."
https://stability.ai/news/stability-ai-secures-significant-new-investmentCan't believe SAI still managed to find new investors after the SD3 fiasco lmao> Stability AI, a leading imaging, language, code and audio generative AI company, today announced it has closed an initial round of investment from a world-class investor group. >he investor group includes top-tier institutions, namely: Greycroft, Coatue Management, Sound Ventures, Lightspeed Venture Partners, O’Shaughnessy Ventures
>>101146920Wow you really moved the goal posts on that one, huh?
>>101146967These investments were secured before the release of SD3 Medium which is why it was released so shit.
>>101146974That's what they're trying to do with the AI model. Commercial use included. Not sneak in a sonic plushie under a "don't care" threshold of SEGA/YT, or for example under fair use applicable only in the USA or only under Vietnamese law.
>>101147028It's absolutely fair use to use a turnaround of a plushie I recorded myself. Which is why you are allowed to upload and monetize it. Also I don't give a shit, society is going to have to realize that copyright and trademarks will have to ONLY apply to outputs. If you produce copyrighted images with an AI model that's your problem just like if you draw copyrighted images.
Anyone seen Ani lately? I've heard some shit is going down and I want to say hi before he loses his job
>>101146813Good luck doing so with something like a Mario. Companies like Nintendo have been one of the biggest pains in the ass of their own community. Same for something like warhammer even. That's why people behind stuff like this want to tread carefully. Training takes a lot of time, effort and big dolla, so better not fumble it. Then again, yes, they can similarly fuble it by castrating their datasets. It's a double-edged sword. In a perfect world, copyright would not exist, but this is not a perfect world.>>101146961Hard to argue with that either.
>>101147135Fair use captures of Mario from recording gameplay video is not the same as using Nintendo Power magazine scans. But we're so far beyond the basic premise, you can have people record things around them and 99.9% of the time it's fair use or straight up legal. The real problem is most of these datasets are scraped images so the right to use these images are largely in question whereas if you had a dataset based on recordings and photos people upload that right to use is much, much clearer.
>>101147081maybe in your place (fair use suggests USA?) this degree of commercial usage is allowed? you misunderstand profoundly if you think this is possible in most of the world, it isn'tof course in the USA you have immense risks because intellectual property law and court cases and lawyering costs and so on are nuts, just stick your neck out with 100ks of trademarks/copyrights/design patents/whatever else and see what happens.before politics or courts haven't categorically decided that AI models indeed do not contain the aforementioned but only learned from them, you're probably in for a serious avalanche of lawsuits. they've started against the other commercial AI too, right?
>>101147367No one is going to train a model based on the podunk standards of your dirt favela.
>>101147397and then you're probably stuck in places that have crazy intellectual property laws and crazy intellectual property law suits like the USAWhat >>101147135 mentioned is only the smallest hint of the craziness there. You better have 100M+ reserved for legal defense kek.
>>101147425US IP laws actually make sense if not slightly overbearing. No one gets sued for IP infringement here and is surprised it happened. Yes, anon, you can't just make and sell your own Mario game. But weird, fan art isn't illegal!
slop everyday
some fucking raucous shitty gens in here today fellas christ
>>101147467Kek good if that makes sense to you. I'll await your model and its successful defense from hundreds of lawsuits including many in fucked up lowest instance courts then.
>>101147670I'll wait for OpenAI's lawsuit then since they clearly have Nintendo IP. :) Fucking favela fag.
>>101146389well that sucks
>>101147699None of it matters, there's at least four major "open" AI models being trained right now. I could've guessed the project Comfy attached himself to was going to be shit.
I CAME!!! AHAHAHAPurple is AdafactorGreen is CAME optimizer, let hope the output isn't full of artifacts
>>101147685In the favela no one cares. It's the USA where you'll probably have to do this dozens of times before it's clear if there's not some small detail that makes everything completely different after all and hundreds of times before you know if states can't do details differently after all.
>>101147740Less loss is good, right? I never quite understood it, but from what little I gather, it roughly means how far off it's from the original input it's being trained on?
>>101147764basically yeah, that's how far from the input training the model managed to replicate
>>101147784then bless the thread in which anon recommended you give it a try, and bless the anon themself
>>1011477640 is a perfect replica. So the closer to 0 the better... until you hit memorization.
>>101147764Indeed, but too fast changing of loss may be caused by artifacts (wrong learning)See, this is the first epoch, there is a lot of artifacts in image.
>>101147801oof, but then again, even for the artifacts, that's a very coherent image, looks promising, and with this being the case, maybe it needs just a bit of adjustment to get really great results
im blessed that forgeUI works on my gtx 960!
god I want to traaaaain now
>>101147847Do you really?
>>101147847get a cheap ARC A770 and be a programmer like me, kek
>>101146316Ai sisters.... We're killing the planet
>>101147874If only there wasn't a very cheap way to produce clean electricity in volume...
>>101147885If only there wasnt a way to store that energy
fun fact, a repeating string of score_schizo can sometimes yeld good/better/interesting results, pic unrelated but I've noticed it because I sometimes forget to avoid putting a style on top of a style>>101147859well, I'm juuust a bit of a masochist I guess>>101147868>sound of non-programmer 8vram laptop tears can be heard dripping the distance
>>101147919You don't need batteries with nuclear.
>>101147874Let it burn. Unlike with crypto, at least we're getting something arguably useful out of it.
>>101147874Same shit as with crypto. Now that megacorps "got mine" and are using it to manipulate information and gamble, the media are attacking "little people" for benefiting from it.
>>101147944Great minds think alike.
>>101147934It's retarded concern trolling that ignores real sources of pollution. Like the amount of pollution that was produced so they can buy yet another purse.
>>101147926Yeah I know but solar etc are the hype du jour right now
>>101147977Yes, surely you own 0 extraneous items designed primarily for pleasure, and are 100% utilitarian and are the utmost conservative in your resting, transportation and living habits
So, we have to make AI green, not just woke?
>>101147977>retarded concern trolling that ignores real sources of pollution.Touché
>>101147998What kind of faggotry is this? I'm pointing out the hypocrisy, I guarantee you the emissions from using my 4090 is negligible compared to the likely hedonistic, rampant consumerism of the author of that article.
>>101148012Don't get me wrong, I'm all for more efficiency and not having to rely on AC and masks in the future. One of the politicians in my country said we should be happy about it, since we'll be able to grow bananas, but I'm not sure if that's how it works, and we're already close enough to being the banana republic of Europe.
>>101148023>I'm pointing out the hypocrisyThe ironic nature of your post is hilarious
>>101147983Build pumped hydro. Further surplus power should be used for waifus and husbandos.
>>101148048I don't concern troll about pollution, faggot. And I certainly don't concern troll about pollution while be a hedonistic normie consoomer. Also kill yourself.
>>101148048They have a point in:>emissions from using my 4090 is negligible compared to the likely hedonistic, rampant consumerism of the author of that article
>>101148045Which country?
>>101148023>its only bad when its something I dont like
Can we please put the discord invite string into the OP of these threads? Thanks all, and peace
>>101148080Consumerism is bad, faggot. No amount of mindless consuming will fill the empty void in your life.
>>101148088No pastebins. No doxcords. Keep drama in other threads.
>>101148065So.. You think they bought the 4090 because of the emissions?
>>101148105Anon you're going to break your back if you bend so far to argue in bad faith.
>>101148096>No amount of mindless consuming will fill the empty void in your life.Dare I say, based?>>101148105No, they probably bought it to train a state of the art coom model.
>>101147983Solar is ass and doesnt generate enough even in desert regions. Wind is a fucking saviour larp too. The more I learn about nuclear the more it becomes clesr that we are shooting ourselves in the foot by not using it.
>>101148113>resorts to insultsClassic
>>101148113>ad hominemFor fuck sake you retard, you're not going to win anything doing this
>>101148132Yeah I think you're a piece of shit when you are intellectually dishonest. >>101148149>samefag or discord raid
>>101148131Wind and solar sounds good for rural areas that are far away from the grid. Otherwise yeah, nuclear is probably THE way to go about it.
>20 drama posts in a row with no gen>Okay, guys, let's keep the drama in the drama thread, okay? Thanks guys
>>101148155>>101148132
>>101148155>discord raidWow, resorting to being a double retard, who would have guessed?
>>101148170>just a coincidence
>>101148163
>>101148183Just keep the discord out your fucking mouth. We kicked you for a reason, fucking nogen debo bitch
Worries about climate. Codes in Python.
>>101148203>just a coincidence
>>101148212Prove itOh, you have no proof?Oh, didn't realise
All these regarded anons in here talking about the discord need to bring receipts next time they're gonna stir shit up because it's getting ridiculous
>the consistent double posting is just a coincidence
>>101147874>cute bird with giant testicles is backblessed
>>101148254>anyone who disagrees with me is all just one person and nobody ever agrees with me because 4chan is so unfaaaiiiirstop whining retard we all hate you
>>101148280how's work today troll fag? still thinking about the shotgun? can't wait to see your centaur spam in sdg tonight
If you're one of the anons calling it the doxxcord, you're the problem.
>>101148294I'm not Ran you imbecile motherfucker so take that name out your mouth before you get in real trouble.
>>101148316prove it with a gen
>>101148328>Someone mentions ran>im NOT ran but shut up or ill doxx youobvious ran is obvious
>sdg dead >massive shitting ITT Can't make this up
>>101148344As far as I know both of the people who got doxxed from the discord did it to themselves with no intervention from the mods, including ran
>>101148368this is divisive nonsensethe problem is schizoanon, whichever thread he has chosen that day suffers
>>101148236>All these regarded anons in here talking about the discord need to bring receipts next time they're gonna stir shit up because it's getting ridiculousIt's a team of schizos. They're figured out how to rattle folks. "Discord" is their new thing. Used to be "pastebin".
>>101148441Well, the discord does actually exist, and one of the mods is the person who runs the pastebin. The constant retard posting about it is done by a small team of /ic/ schizos, though, I'm sure of that.
>>101148298>doxxcordThey want your email. Case closed.
i miss debs, bros, not been any news posts in a while now
>he still doesn't know what's about to happen
>>101148488your workflows too
>>101146698its not artists, it never was, or rather they were dragged into it by false flags. its corporations who dont want the fancy new tech to be open source or free.so of course they dont care and only want it gone.
>>101148658No, there are real artists being vocal about this. They are threatened by the loss of money and they basically say anything and everything to get AI banned. These are the people screaming in every social media venue.Corporations are afraid of being sued and a lot of the peons at these companies are also trying to be the ethical and moral police, I have no doubt many actions taken by OpenAI and SAI were done by people who are trying to force people to "do the right thing".
>>101148744>No, there are real artists being vocal about thisfor sure they are, but theyre being duped or are paid to do it by those corpos.i remember some stupid "copyright alliance" where they tried to get aspiring online artists riled up to team up with disney and the like.
>>101148369Ran is not a doxxcord mod
>>101148799Anon, there doesn't need to be a conspiracy. Luddites exist. Any drooling retard who can't abstract think will think AI is bad as a gut reaction. These people are always there screaming, they did the same thing when digital art was becoming popular. The tradition media artists were shitting and pissing everywhere because digital art made art cheaper to get into.
>>101148836of course, but there is also a conspiracy.
a lot of great gens today guys, nice work
>>101148856No, social media just makes it easier than ever for retarded bandwagons to be created.
Discord has been popping off lately if anyone's looking to join
>>101148938
>>101148886Sorry boss, been out a few days. Will get back to work
>>101148471> /ic/ schizosThis explains the tree thing, but it shows how disconnected you are. The issue is the avatars themselves, they are the ones doing all of this. Making up a boogeyman because you don't want to face the real problem is only going to make things worse.
>>101148813Uh, yep, he is
>>101148964How do you deal with them? They are very active, bake most threads and have by far the majority of posts, both gen and nogen. What is puny tiny little you going to do about it?
>>101149001If you close your eyes it's like they don't exist and they don't matter.
>>101148899which in turn makes it easier for corporate astroturfing. of course these big media publishers dont want you producing your own media, and of course big tech companies want to keep this to themselves and to sell on their terms, theyre already selling it.
>>101149001just ignore them dude
>>101149019They have completely consumed this thread. There is almost nothing here but their ramblings
>>101149036What Disney does or does not want with AI models is irrelevant to me. SAI's SD3 was ass not because of Disney but because of handwringing puritans.
>Debo>Ani>Ran
>>101149001report them for off topic, and posts mentioning the discord too. they can go back to their thread
>>101149051i think there would not be hardly as much opposition if people were wise to their tricks. they spread lots of disinformation.
>>101149057Pick two
>>101149085The handwringing is self inflicted, SD3 wasn't ass because it contemporary artists.
>>101149001thinking about moving to another board, the problem is, which other blue board allows local imggen?
>>101149132i was replying to a post about angry artists, not about sd3s quality or lack of.
>>101149152Several of them allow DALL-E, but you cannot possibly make a new thread without them immediately jumping on it to kill it in the womb. The first 3 posts to your new thread will be "Debo" "Fuck Ani" and "Schizo anon is sooo back"
Isn’t it time to merge with the main thread? United we’re stronger, divided, we’re weak.
>>101149174Angry artists don't need corporate agitators to exist.
>>101149190This thread has a Chinese model focus and the other thread has an SAI focus. There really isn't much overlap in the content discussed.
>main thread
>>101149190no this is what one particular anon wants. the same one trolling this thread.
>>101149208D E B OEBO
>>101149206You could just talk about both in the same thread.
>>101149206stop making shit upwe just dont want avatarfags, doxx discords and dramaevery local image generation model is welcome here
>>101149208>every post i dont like is one anon>the thread would be great if it was only me>>101149238There is basically no overlap between the two topics
>>101149205i dont want to go in circles, sorry.
>>101149251It’s the same topic, and the same thread.
>>101149241This is the main thread that the discord members/avatarfags post in
>>101147859>571 daysdamn>>101149206>This thread has a Chinese model focus and the other thread has an SAI focusnonsense. Local models go here
>>101149265It's two completely different topics and the threads share very little in common./ldg/ has a lot more gens per post and a lot more serious discussion./sdg/ has very few gens comparatively and a lot more of this retard schizo shit.
>>1011492800/10 for the effort, better luck tomorrow
>>101149298Discord mod hands wrote this post
>>101149292And together you would bring the best of both worlds. This kind of division is going to kill open source.
nb4 a joint thread is baked only for sdg to immediately double bake like last time
>>101149314Fuck unity, I'm literally not going to spend more than 2 more minutes of my life thinking about SAI slop like SD3. I'm interested strictly in the future models
Reminder not to bother with ragebait shitposting. Spare yourself the brainrot by refusing to engage with it. There's better things to do, and more worthwhile discussion to be had.
>>101149339/lsdg/ when
>>101149314Just let it go man
Sisyphus rolling the god damn boulder again
every launches ok with stock a1111 with the amd gpu use set up, i have low vram on but i'm on 8gb of vram, and as soon as i try to generate, it hangs. gives me 0 error after a few minutes, and then press any key to continue then quits. any ideas?
>>101149415Loving every it/s of it.
>>101149415confirmed sissy baka
>>101149282Hard to know what the real time is. Might just be 2 or 3 epochs, it's training on around 6 million images now.
>>101149495do the samples look like anything yet?
>>101149573
>>101149415I love you anon
>>101149590Looks like a transformers screaming "god left me unfinished".
>>101146316https://chipsandcheese.com/2024/06/25/testing-amds-giant-mi300x/>finally some good fucking food
>>101149590little bit rough, but I guess it's getting there
>>101149495>6 million imagesthat's a lot for a finetune, did you make some other models before that one?
>>101149788It's a from scratch model, 1.3B parameters made on Pixart architecture.>>101149751It'll get there, it has to learn from ground zero, it's only effectively 200 hours of training so far.
>>101149668>>101149694These don't look like /ldg/ gens to me. Intruder?
>>101149816>It's a from scratch model, 1.3B parameters made on Pixart architecture.you're continuing the pretraining of pixart? you know that they will release a more advanced base pixart model? maybe you should wait for their definitive version before working on it aswell, that's my 2 cents
>>101149871No, it's something I'm doing from uninitialized weights I defined, this model is training from literal zero. Pixart Sigma is 0.6B. I don't know when they're releasing a new model, if ever and it took them months between Alpha and Sigma. Their next model, if ever released, will be an upgrade / new model just like Alpha to Sigma. For now it's my hedge as a different open model, it's training on an uncensored dataset, and I'm doing it as a hobby project.
>>101149824>These don't look like /ldg/ gens to me. Intruder?I'm out of the loop here, what?
>>101149913I see, that's a very interesting project you decided you tackle your hands on, I wish you all the success because that would mean we would get a fully uncensored base model ready to be used. For the captions what did you use? Laion's captions or something like CogVLM?
>>101149913Don't let your memes be dream anon.
>>101149962Honestly I think it's learning faster than I could have hoped given the numbers they gave in the paper.
>>101149991That and the fact you CAME really shows promise, even if within it's paramter's limits.
>>101150009I don't understand what you're saying. But honestly I'm just hopeful because it means Pixart's architecture might mean we're eating good forever. If I can do 1.3B, I can do 2B and beyond especially as hardware improves. We might very well be training a 8B model 10 years from now on Pixart using consumer hardware (8090s).
>>101150036what hardware are you using to pretrain a pixart 1.3b?
>>101150036Aren't you using the CAME optimization instead of Adafactor, or do we have two anons training and I got confused.
>>101150073I'm training on two 4090s>>101150078No I'm a different anon, I'm doing AdamW 8-bit which gets me batch size 32 at 256. So it should scale okay to 1024px.
>we have TWO anons trainingreal /ldg/ hours
>>101150101>I'm doing AdamW 8-bitwhat do you think about this new paper? it says it makes the training even better or some shithttps://arxiv.org/abs/2406.16793
>>101150146I'll look into it, anything that makes memory go down is good. I tried 4-bit training too but it was unstable.
>>101150152>swordlanceokay, I'm intrigued
>>101149348what's the lora for this one?
>>101150389https://civitai.com/models/531020/wakfu-anime-style-poneyxl
>>101150667thanks
>>101150913very punk>>101150973very arty
>>101150169Sadly it's way too fat compared to 8bit AdamW. Couldn't even do batch size 4 at 256.
>>101150973c-catbox?
>>101151367https://files.catbox.moe/zhg680.png
>>101151367also if you want the lora:https://litter.catbox.moe/xy9d2z.safetensorsas mentioned earlier it's a failed gyategyate training
>>101151706>>101151791thank you anon
>>101151861
>>101151284JK got it working with bf16
>>101151806np. but it really just is a failed training, so most people wouldn't wanna bother I think
>>101146389As long as it's curated high quality images it doesn't matter.
>>101152231You actually believe they're going to curate 10+ million AI slop images? They don't even look at the real images they train on.
>>101146389Not my loss if they try, but yeah this is almost certainly dead on arrival.
>>101152246You can automate most of it if you bothered keeping up with current training methods.
>>101152297>automate making sure the AI images even make senseHere's the thing about real images, I can trust people don't have an arm coming out of their knee.
>>101152246You can do this with https://github.com/chaofengc/IQA-PyTorch/ for example.Use qalign and some others as well as popular images and you probably have pretty nice images.
>>101152318Have you considered one of the reasons why SAI products look like shit is they overuse automated "this image must look good based on math" filters?
>>101152326No I have not. People just want to make shit up.Besides you're not finished in one iteration.
For stylized it's not bad at all. https://huggingface.co/datasets/ProGamerGov/synthetic-dataset-1m-dalle3-high-quality-captions
>>101152326SAI are shit at it is the issue. They used shitty cogvm. There are far superior options now.
>>101152345Make what shit up? That SD overproduces bokeh images?>>101152353Training on this will guarantee mangled hands.
>>101152354Very nice
>>101152377>Training on this will guarantee mangled handsHave you seen Laions datasets? Curated ai generated images are better than most of that garbage.
>>101152407???Anon, you need the junk. You need random eBay images. Your dataset can't just be a bunch of AI generated Reddit memes. And you have to remember, these are the images they cherry picked. Dalle is more than capable of generating absolutely piles of shit. I'll take someone's random MySpace pictures over AI generated slop.
>>101152213sometimes the failures produce the most interesting results
>>101152377Basically all of what you said, that > SAI products look like shit is they overuse automated "this image must look good based on math" filters?That is simply not a fact, the aesthetics and image quality scoring algorithms largely concur with human assessment anyhow.
>>101146534this
>>101152471>mostlyUh huh. Good cope. And what a coincidence, it just does really bland stock images with excessive bokeh. Almost like it's aligned with what is mathematically "appealing".
>>101152484yea most humans even like the bokeh AND you can literally just do a model that prefers not having the bokeh or out of focus areas if you're with the other group.
>>101152502My point is if you use aesthetic scoring you're going to bias your model and I'm going to laugh so hard when the AI slop model with 10/10 aesthetic scored images comes out looking like shit and everyone wonders what went wrong.I'll tell you right now that aesthetic scoring is not great, at all. Lots of false negatives, lots of false positives. And frankly, the goal is not to eliminate the bad images, but to properly annotate them as bad.
>>101147859Hype! Grad norm is pretty high though anon, might want to get that checked out.>>101149590Hype!!! 256x256 of pure HYPE>>101152502>most humans even like the bokehTo add, the more/better bokeh, the more expensive the camera lens. Aperture can only go so wide w/o sacrificing quality
>>101152528Humans very much agree with modern aesthetics scoring.As for technical image quality assessment (e.g. was this compressed heavily or lightly, did a better camera record this or a worse one), AI does that even better than random humans.Yes, you probably should not completely eliminate the worse data when training AI also because perhaps you get all 3DCG and no 1990s digital camera photos? But biasing the training towards aesthetics is just fine.
>>101152600Like I said, you'll just get SD3 and SDXL which have zero range. You just sound like someone who doesn't know contrast from hue.
>>101152613Like I said, you have nothing. SD3/SDXL are literally not showing anything one way or the other. It's literally nothing.
>>101152742You simply confirm you lack taste. Have fun with your DE3 aesthetics model.
>>101152753You realize how stupid it is to claim this results in DE3 aesthetics JUST after claiming it results in SD3/SDXL with their different aeshetics?It does not result in any specific aesthetics tho I suppose you get attractive 1girl/1boy on average since humans and AI are trained to rate that highly for aesthetics.In the end we hit flaws in the technology anyhow, you can train data but it doesn't mean that combinations of tokens always will get a great score AND high adherence to the concepts they represent in a sensible combination for the tokens presented.
remember when these threads were fun?
>>101152849lol you're funny, at least you make up in wit what you lack in taste
>>101152860didn't someone always make asspulls?
>>101152171nice anon
Hello anon
>>101146316nice collage desu
>>101153786Howdy stranger
Kaboom
>>101154012
>>101154012>me after having indian food
>>101153969yee haw partner
Have you prepared anon?
Good night anon
I miss chang
>>101154635gn kot
>>101154190Minima?
>>101147874I'm pretty sure that the power grid is being exhausted by giant data centers and GPU warehouses, not by anon generating anime titties. Image models are pretty small and no more power consuming than your average PC game
>>101152860I like learning
>>101149206what Chinese model
choose:>dead thread>pedo posters, pedo apologists, a single anon masquerading as other anons/avatars
>>101155225Second one
Nice hours >>101150683Nice
worthless dead general the only posts are made by ran samefagging and chink bots lol
anyway
>>101155037there are like three of them
What's the best model for animeslop?
>>101155726Pony/t-ponyai3
The art gallery has opened:>>101155934>>101155934>>101155934