Previous /sdg/ thread : >>101562401>Beginner UI local installEasyDiffusion: https://easydiffusion.github.ioFooocus: https://github.com/lllyasviel/fooocusMetastable: https://metastable.studio>Local installAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiComfyUI: https://github.com/comfyanonymous/ComfyUISD.Next: https://github.com/vladmandic/automaticAMD GPU: https://rentry.org/sdg-link#amd-gpuIntel GPU: https://rentry.org/sdg-link#intel-gpu>Use a VAE if your images look washed outhttps://rentry.org/sdvae>Run cloud hosted instancehttps://rentry.org/sdg-link#run-cloud-hosted-instance>SD3 info & downloadhttps://rentry.org/sdg-link#sd3>Try online without registrationsd3: https://huggingface.co/spaces/stabilityai/stable-diffusion-3-mediumtxt2img: https://www.mage.spaceimg2img: https://huggingface.co/spaces/huggingface/diffuse-the-rest>Models, LoRAs & upscalinghttps://civitai.comhttps://huggingface.cohttps://aitracker.arthttps://openmodeldb.info>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>Share image prompt info4chan removes prompt info from images, share them with the following guide/site...https://rentry.org/hdgcbhttps://catbox.moe>Discord6wUwtcJsr2>Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
Hello
>mfw Resource news07/25/2024>ViPer: Visual Personalization of Generative Models via Individual Preference Learninghttps://viper.epfl.ch>ComfyUI-Kolors-Translator: Translate prompts into Chinesehttps://github.com/BetaDoggo/ComfyUI-Kolors-Translator>ComfyUI-FollowYourEmojiWrapperhttps://github.com/kijai/ComfyUI-FollowYourEmojiWrapper07/24/2024>SV4D: Dynamic 3D Content Generation with Multi-Frame and Multi-View Consistencyhttps://sv4d.github.io>Open-Sora-Plan Report v1.2.0https://github.com/PKU-YuanGroup/Open-Sora-Plan/blob/main/docs/Report-v1.2.0.md>Official global launch of Kling AI's International Version 1.0https://klingai.com>INF-LLaVA: Dual-perspective Perception for High-Resolution Multimodal Large Language Modelhttps://github.com/WeihuangLin/INF-LLaVA>FoRA: Low-Rank Adaptation Model beyond Multimodal Siamese Networkhttps://github.com/zyszxhy/FoRA07/23/2024>Mammoth - An Extendible (General) Continual Learning Framework for Pytorchhttps://github.com/aimagelab/mammoth>Differentiable Convex Polyhedra Optimization from Multi-view Images https://github.com/kimren227/DiffConvex>Cinemo: Consistent and Controllable Image Animation with Motion Diffusion Modelshttps://maxin-cn.github.io/cinemo_project>Automatic Generation of Fashion Images using Prompting in Generative Machine Learning Modelshttps://github.com/georgiarg/AutoFashion07/22/2024>Intel AI Playground beta Official Launchhttps://game.intel.com/us/stories/introducing-ai-playground/>SD3 Unbanned: Community Decision on Its Future at Civitaihttps://civitai.com/articles/6221/sd3-unbanned-community-decision-on-its-future-at-civitai>GenTron: Diffusion Transformers for Image and Video Generationhttps://github.com/lavinal712/GenTron>ComfyUI-Lumina-Next-SFT-DiffusersWrapperhttps://github.com/Excidos/ComfyUI-Lumina-Next-SFT-DiffusersWrapper>PhotoMaker V2 Released with improved ID fidelityhttps://github.com/TencentARC/PhotoMaker#-new-featuresupdates
>>101574948New Update to Lora:X-Men Evolution Style SDXL V2:https://civitai.com/models/578582v2 changes: Much better faces and cleaner outputs.Has character prompts now (if you want to force a certain character. Or have their influence)
>mfw Research news07/25/2024>XMeCap: Meme Caption Generation with Sub-Image Adaptabilityhttps://arxiv.org/abs/2407.17152>Looking at Model Debiasing through the Lens of Anomaly Detectionhttps://arxiv.org/abs/2407.17449>HumanVid: Demystifying Training Data for Camera-controllable Human Image Animationhttps://arxiv.org/abs/2407.17438>Deep Spherical Superpixelshttps://arxiv.org/abs/2407.17354>Revolutionizing Text-to-Image Retrieval as Autoregressive Token-to-Voken Generationhttps://arxiv.org/abs/2407.17274>LPGen: Enhancing High-Fidelity Landscape Painting Generation through Diffusion Modelhttps://arxiv.org/abs/2407.17229>MemBench: Memorized Image Trigger Prompt Dataset for Diffusion Modelshttps://arxiv.org/abs/2407.17095>When Text and Images Don't Mix: Bias-Correcting Language-Image Similarity Scores for Anomaly Detectionhttps://arxiv.org/abs/2407.17083>Q-Ground: Image Quality Grounding with Large Multi-modality Modelshttps://arxiv.org/abs/2407.17035>Diffree: Text-Guided Shape Free Object Inpainting with Diffusion Modelhttps://arxiv.org/abs/2407.1698207/24/2024>PartGLEE: A Foundation Model for Recognizing and Parsing Any Objectshttps://provencestar.github.io/PartGLEE-Vision>MovieDreamer: Hierarchical Generation for Coherent Long Visual Sequenceshttps://aim-uofa.github.io/MovieDreamer>DreamVTON: Customizing 3D Virtual Try-on with Personalized Diffusion Modelshttps://arxiv.org/abs/2407.16511>Harmonizing Visual Text Comprehension and Generationhttps://arxiv.org/abs/2407.16364>DreamDissector: Learning Disentangled Text-to-3D Generation from 2D Diffusion Priorshttps://chester256.github.io/dreamdissector>Outfit Anyone: Ultra-high quality virtual try-on for Any Clothing and Any Personhttps://humanaigc.github.io/outfit-anyone>Channel-Partitioned Windowed Attention And Frequency Learning for Single Image Super-Resolutionhttps://arxiv.org/abs/2407.16232
>>101575014nice, how to get the clothing?
>>101574959howdy>>101575000do you people actually read all these papers? I usually skim the abstracts but even then, that feels overwhelming
>>101574948Ah, one of mine again. I guess they make good OP images because they're... plain, but cute.
>>101575260there can't be more than a few people who actually read the news, especially when you consider who's posting them
>>101575250She is carrying your nuts.
>>101575260no, I read enough to make a determination if it is relevant to imggen or might be interesting to people in the general, then I throw it into the pile or throw it in the trash. I'm pretty much too dumb to understand most of the technical stuff anyway; I'm just trying to play my part in circulating knowledge and sparking discussion>>101575439people are for sure much more interested in the resource news cuz thats where all the new toys and genAI drama shows up. but there's still people who like the research stuff. even for people who don't read it, its motivating/inspiring to see how much stuff is happening in the field
>>101575467tldr
>>101575024"woman standing in a garden, white soft lace bandeau, short cloth around her waist"Still not *quite* what I was going for anyway. I was envisioning a "skirt" that looked more like a bandeau, but with a lace trim. I'm not even sure how to go about that
>>101575470exactly
>>101575467>sparking discussionCan you link me to the last discussion about something in the news and anything relevant anon found?
>>101575480thanks
>>101575467>I'm just trying to play my part in circulating knowledge and sparking discussionyou are doing good work anon, I appreciate you
Believe it or not, the Michael Jackson tipping was accidental. All I did was>1girl, a girl wearing a bikini, dancing at the beach, long hair, digitigrade legs
>>101575544Nothing is an accident in latent realms.
>>101575503its never happened yet but I'm hopeful today is the day>>101575521thanks!
>>101575558When's the last time it happened?
>>101575573ask with a gen and I'll give you an answer :3
>>101575521any gens to share?
>>101575634Why?
>you MUST identify yourself when engaging with debo
https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_dit/I implemented hunyuan DiT and properly packaged the model into single easy to use checkpoint files.
Newfag here, now Tensor.art is giving me exceptions whenever I try to generate anything using the same workflow that was working yesterday and I'm too retarded to figure out how to fix whatever happened because I can't even find a log of what went wrong. Is that a common thing on that site?
>>101575680based. any spoilers on whats up next in the queue?
>>101575726Can you answer >>101575573 >>101575645 ?
Why do you refuse to answer, Debo?
>Hello anons! I hope everyone is doing well :]
>>101575503malware
>>101575807kek
>>101575807That comfyui data logger thing?
>>101575819no it was debo including malware in the news
>>101575819no, actual malware.the guy asked for their node to be included "just for fun"
>>101575826Okay, what? I admit I haven't paid much attention to the "news". Is there a link or something?
>>101574948anime is gay you fucking faggot. embrace your own cultures. god fuck japanese yellow nigger shit. embrace Asgard and Olympus. embrace European western culture.
I made a bmp>>101575826if you scrolled past the news without reading it, you got malware. I still think it was justified
https://files.catbox.moe/u7qksw.gif
>>101575849https://desuarchive.org/g/thread/100878638/#q100879769
>>101575854https://desuarchive.org/g/thread/100878638/#q100883674>first chance to say "ah my bad" >blames anon instead
>>101575849>Is there a link or something?one of the hundreds of comfy nodes I included in the news started installing malware on people's machines. the malware was added after I featured it in the news but some people pretend that I somehow knew the node would have malware in it and purposefully spread it.
>>101575923>people pretend that I somehow knew the node would have malware in it Link to anon saying that?
Anyone ever have 4chan refuse an image for seemingly no reason? Doesn't give an embedded file message or anything. You click post and the image doesn't go through and you get no message and if any new posts have been made you auto refresh and the new posts appear but the reply window remains open without posting the image. Ever had this happen? It's happening to me right now with two specific images. Even tried changing their resolution and encoding to .jpg. Bizarre. They can't be banned images because I literally just made them.
>>101575939I've seen other anons say thats happened to them. I haven't experienced it personally
>>101575981Any links for >>101575938 ?
>>101575939I've had a thing where I click Post and see the upload percentage...then the button changes back to "Post". Click Post again and it says I got the captcha wrong
>>101575994Yeah that's what's happening.
>>101575680Good work, Chang!
>>101576059Wow. Aggressive cropping didn't work. Changing to .jpg didn't work. Saving it with a completely different application didn't work. Clearing cookies didn't work. Changing IP didn't work.Adding transparent borders to the sides got it to post.Fucking bizarre.Well that was a fun little experiment.
>>101576068Talk about being obsessed jesus christ
>>101576119Doesn't Debo call you "Troony"?
Test
>>101576328>hand clipping through hair ornament
>>101576307Diffusion is /g/
>>101576393anime is not /g/>>>/a/
>>101576407Diffusion is not /a/, we'd get sent back to /g/
Dammit, my IP range is banned on /v/
>>101575680what's next? Kolors? AuraFlow?
>>101576428You're not missing much. Been a shit hole over there for a long time.
>>101576411he got filtered by the webui install and now he's spending his time lashing out instead of being productive
>Automatic1111>Masked in-painting in img2img using ADetailer only not workinghttps://github.com/Bing-su/adetailer/commit/09221b1065d4ed889c0fab485f6358909acf404bMy understanding is that it's now disabled for some reason? Why? Do I just have to now manually crop my image into chunks and run ADetailer on the individual chunks without in-painting?
>>101576476Do you have any links for >>101575923 ?>some people pretend that I somehow knew the node would have malware in it
>>101576449It's a diffusion output. It's /g/.Go back.>>>/b/
>>101571395Did not forget about you. I think she's picrel that you mentioned, but here's my upscales from this past week togetherhttps://files.catbox.moe/hppexo.jpghttps://files.catbox.moe/pivld5.jpghttps://files.catbox.moe/hrpk8p.jpghttps://files.catbox.moe/ph4fba.jpghttps://files.catbox.moe/jjjbdd.jpghttps://files.catbox.moe/gw43i7.jpghttps://files.catbox.moe/2l3ae1.jpghttps://files.catbox.moe/5j5kc8.jpghttps://files.catbox.moe/8s2fu9.jpghttps://files.catbox.moe/66k0zg.jpghttps://files.catbox.moe/ee293k.jpghttps://files.catbox.moe/dt1ztc.jpghttps://files.catbox.moe/mbkwm6.jpghttps://files.catbox.moe/gsd0qc.jpghttps://files.catbox.moe/4e9ziw.jpghttps://files.catbox.moe/tmb9zl.jpghttps://files.catbox.moe/rh7938.jpghttps://files.catbox.moe/o5iz19.jpg
>https://civitai.com/models/4514?modelVersionId=5119>[:(detailed face:1.2):0.2]: When drawing high-quality faces, do not use the "detail face" tag at 0 steps, otherwise it may lead to deviation from the original semantics of embeddingCan someone explain this to a noob? What does that detailed face formatting mean? :1.2 is the weight but ():0.2 on top of that?
>>101576731No links for anon?
>>101576688>[:(detailed face:1.2):0.2]this is a prompt editing sequence[one:two:step]'one' and 'two' are tokens. 'step' is the step when 'one' changes to 'two'. steps can either be an integer for a specific step or a decimal for a percentage of total stepsso back to your prompt. 'one' is an empty string, 'two' is (detailed face:1.2), and steps is 20% of the total steps. that means for the first 20%, there is no token, then from the other 80%, there is (detailed face:1.2)>ok but why?thats a good question. I guess the author found that (detailed face:1.2) causes distortion in the first 20% of convergence? it seems rather arbitrary but I guess he knows something we don't
[one:two:step]
>>101576781Any links to back up your claim in >>101575923 ?
>>101576781Thank you for the explanation. Now that got me thinking since I didn't know that was possible before. Let's say there's an object lora I like but it tends to fuck with the style of the gen too much and lowering the weight tends to ruin the results specific to the lora. Would that sort of one two step approach possibly help in fixing that?
>>101576851you can't use prompt editing to add/remove a lora because the lora effectively gets appended to the model before sampling. if the model was trained with an activation word, maybe you get some toggling control? I've never been exactly clear on how that works tbdesu
>>101576781Chronic idiot strikes again
>>101576912Do you have any links to the claim that anons said you knew there was malware in the news >>101575923 ? Also, how do you feel to only have uninformed newfags take your advice?
>>101576912It has an activation word. I guess I'll have to play around with it to see what effect it has on the end result.
do you guys activate your cuda stream ?it says it might speed up the genning but it feels slower when my almonds are activated
>>101576978yes, I do morning yoga with my GPU to activate my cuda chakras. hte cuda flow definitely makes a big difference in bot how efficient it runs as well as how creative the gens are. I'll never go back to unactivated cuda
>>101577008Did you just make a false claim in >>101575923 or do you have any links to back it up?
>>101576648THANK YOU!
>>101577008>I'll never go back to unactivated cudaAlrighty, thanks breI do feel like the quality improved but hard to say
>>101576688>>101576978Any gens, anon?
>>101577050Can't post them on blue board.
>>101577050im >>101576978sadly at work + they are very nsfw :s
>>101577069>>101577071You can always upload to catbox and post a link here and jannies wont care.
>>101577069>>101577071
>>101575923>>101577017https://old.reddit.com/r/comfyui/comments/1dbls5n/psa_if_youve_used_the_comfyui_llmvision_node_from/
>>101577076I don't have catbox.>>101577080I'm not a furfag or a p*do. Not sure what else you might be implying.
>>101577127>Not sure what else you might be implyingoh nothing that sinister, just coomers kek
Any style loras that are similar to ma-ko's style? There doesn't seem to be one for his style specifically.
>>101577134Oh I read too deep into it. Yes I must admit I am a coomer. And for now I'm a pleb proompting off buzz on civitai so I must carefully choose what to spend it on. Coomslop tends to win out. I used to do sfw proompting on bing but they restricted it too much and took all the fun out of it.
>>101577159Coomers? On 4Chan? Think of the advertisers!
>>101577243nice
>>101577256Inspired by the bikini gens above I whipped up something I can post here.>>101577243Absolutely delicious butt.
>>101577296>>101577300
>>101577243>>101577349what is this style?
>>101577375https://civitai.com/models/471461/did-vaporwave-visions-pony-xl
gn sdg
>>101577592Who would've guessed this...
A new paper suggests that training an AI with AI pictures is a really bad ideahttps://www.nature.com/articles/s41586-024-07566-y
>>101577617Hey debo, remember when you asked for proof that training a model with synthetic data is retarded? Well here it is.
>>101577617baker, add this paper to the op, more morons need to see this
>>101577673Right away, Sir! I am sorry I didn't reply immediately. Won't happen again.
>>101577682i'll be turning your bread into cream buns the next time you fuck this up, understand?
Grr...
>>101575461catboxu?
https://reddit.com/r/StableDiffusion/comments/1ecg544/a_girl_emerging_from_the_hot_springs_adjusting/https://reddit.com/r/StableDiffusion/comments/1ecfu3g/girl_turning_around_for_the_camera_showing_her/bro, hentai is gonna be amazing in the next few years
i miss schizo anon
>>101577617if you trained on shit irl images you'd get basically the same. curation and filtering is a thing
>>101577894read the paper, it's not that simple, no matter how good the filtering is, it will act as a poison to your AI anyway
>>101577617Can't believe we need a fucking paper to convince people that training an AI with synthetic data (that garbage data has an inherently inaccurate anatomy/perspective/illumination) is a dumb idea.
>>101577915big booba
>>101577914> We find that indiscriminate use of model-generated content>indiscriminate the paper makes no mention of filtering or curation. it's yet another instance of researchers stating the obvious
>>101577617does that mean AGI is DOA?any recursion would be a downward slope
>>101577941There seem to be a discution about that on twitterhttps://xcancel.com/eirshy/status/1816588706431795525#m
>>101577945>does that mean AGI is DOA?if you use synthetic data to train your model yeah it's DOA, but if you go for real picture, the limit that can be achieve will always be reality
>>101577939And gun
>>101577965>SD3 blocks your pathshit in, shit out. simple as
>>101577957>xcancelReplacement for nitter? Thanks.
>>101577979yeah it's the best nitter-like server I found, It works great
>>101577617>It's 2024 and we discovered that recording a VHS on a VHS makes the result worse than the originalhttps://www.youtube.com/watch?v=nqy_hYDI0As&list=LL&index=3&ab_channel=JaphyRiddle
>>101578008like using a key cutting machine on each new copy
>>101577760Only when we get local versions
>>101574948>Press any key to continue . . .keep getting this bug when I try to gen something
>>101578031The best local model we have so far is OpenSora, it's shit and asks for 67gb of VRAM, there's still a ton of work to do to get there...
Were there any new good anime models in the last 2 months or so? The last time i checked it was all pony autism mix.
>>101578041Jesus
Coomer here, say I wanted to make stuff like this but with my Skyrim characters. How many pics of them would it take to train the model? Hundreds, thousands, tens of thousands?Also, I bet I already kow the answer, but can this shit be done on an AMD card at all?
>>101577617And yet all """open source models""" work by taking Llama, then automatically training it on GPT outputs.This is why I hope base Llama improves a lot; I'm not going to trust Shitstral to do more than finetune it on GPT slop with a particular focus on questions known to appear on benchmarks.
169m, from only ~2.5 months7m, ill check dates to be sure but pagination stopped on one node, must be all they have350k actorsstill no downloads on 274m set, i guess filtering is too difficulti thought there would be interest in doing large scale comparisons of sampler, cfg scale and everything else, maybe there's a reason some images suck in the first place. or idk detecting anomalies and fixing them
It's fun to look through old images and pick up old prompts. Seems like a big regression happened during this year or something. Brainrot I guess.
>>101578212>Seems like a big regression happened during this year or something.I think it's genuinely just >>101577617
>>101578232I was referring to my own enthusiasm and interests... old setups are cleaner and more organized, prompts are more interesting and so on.
>>101578091>Also, I bet I already kow the answer, but can this shit be done on an AMD card at all?check the op, there's a link to some guides for AMD gpu. Apparently it's better under linux but can still work on windows.
>>101578232with the exception of fal nobody is indiscriminately training on ai outputs, even then, the general reception is that the model is ok and they just needed to filter out the fat cat blocked generation images. look at lexica and playground, theyve finetuned on their own outputs but with curation, its created a unique style that people love and all they had to do was allow users to indicate whether they like an image. even if there's not a like button the good ones are tracking which generation out of a batch that you download, that's why they give you 4 images at a time
>>101575680Thank you for your work, do you plan on implementing native pixart support?
>>101578337it's still a retarded idea to train an AI with pictures that have innacuracies when real pictures that are perfectly accurate to reality exist
>>101578461thanks, captain obvious
>>101578504>captain obviousso "obvious" everyone is still training their models with synthetic data, looks like the world is surrounded by retards yeah
>>101578511>train on inaccurate images>get inaccurate resultshmm.. it must be all synthetic data that's the problem, can't possibly remove the inaccurate images, it's not like training on real world inaccurate images produces inaccurate results
>>101577651Did he really say that?
>>101578548>it's not like training on real world inaccurate images produces inaccurate resultsthe fuck does that mean? a photography is always accurate to reality, unlike AI that will always have some mistakes on lightning, anatomy or perspective
>>101578565>mistakes on lightning, anatomy or perspectiveright, i forgot, these mistakes never exist irl
>>101578573how can they exist, a photograph is always accurate to reality, the fuck anon?
>>101578573photos are always true to realitythis doesn't apply to shitty r34 art tho
>>101578611>>101578618yes of course, it's impossible to take a bad photograph, every real world image is extremely high quality
>>101578573nta but image gen is not at the point where it can compete with real images and art, there's always little flaws that appear in ai art. there is a reason why ai slop appears so uncanny while real art and images (even if they are bad) do not. the human brain prefers the flaws of reality (photos) and human mistakes (art) over ai shit.
>>101578635double down, kingmaybe this way reality will bend to your will
>>101578635that's ironic because an AI needs diversity of data to be great, so having bad photos and telling the AI it's bad photos is required to get some good training, especially if you want to put "bad quality" on the negative prompt. But if you add AI pictures you tell the model this is the perfection it should achieve when it's not, then you're being a retarded nigger that doesn't know shit about life, I really hope you're not in the 2nd option anon
>>101578151>>101578212Nice castles
>>101578670you made the animation with kling?
>newsfags trying to have an honest discussion with thread schizonever gets old
eu hours are so dumb
>>101578675No, using Kong instead.
>>101578767you should try kling, you can get 6 free "image to video" process per dayhttps://youtu.be/EIj9-xgfV2c?t=70https://klingai.com/
>>101578696thread schizo is asleep but will be up in maybe 3 hours this was his last post >>101577008
>hans, ivan and pierre trying to figure out why their model still sucksjust 2 more billion images
>>101578799Mon ami...
>>101578794but we're all the thread schizo
>>101578854we are all schizo anon but there is only one thread schizo
>>101578874kek this,G'mornin Anons,
>mfw posting in the debo containment thread
>>101578890>eu hours are so dumb
>>101578675it's img2vid with animatediff, which comes with infinite free processes per day, and I finally stopped being lazy and set up masks so I can define the parts of the image that I want to animate. No more morphing buildings and mountains.Btw is it possible in ComfyUI to set the behavior of the ctrl + scroll mousewheel combo in the mask editor to zoom on the mouse location instead of the top left corner of the image?
>>101579014that looks really good, good job anon
>>101578794sure thing "anon"
>>101579014Nice work.
is this the onegirl thread?
>>101579373That's right!Welcome!
>>101579029>>101579167Thanks anons
>>101579373is there some saoirse in that prompt somewhere?
>>101579393What is she scheming?
>>101577617we needed a research paper to realize this?
>>101579014anything special about looping or just a masked i2v workflow?
>>101579421good eyebeen a while since I've done this, hard to find gens with acceptable hands. Also need to learn how to upscale properly again so I don't get extra torsos
>>101579431To turn this thread into 1girl anime central!
Good morning
>>101579543this is lovely, ty
>>101579585Don't touch it... it's mine!
>>101579437Looping is set with the Context Options - Looped Uniform node. Context length is 16, overlap is 2, and the animation is 28 frames. Otherwise it is a standard animation workflow (with ipadapter, depth controlnet, and freeu). I couldn't get it to work properly until I used the InpaintModelConditioning node and Multival Scaled Mask.
https://suno.com/song/a1573828-cc04-4faf-aa8f-64e4053e35afsome git metal>You seem to have mistyped the captcha.Your robot sucks, actually.>KRPRkek
>>101579513good morningi hate trani
>>101579697thanks, i will probably get into 1.5 again just for animatediff
>>101579827I'd suggest one of the humu models in case you haven't tried them. both v1 and v2 seem to get the cleanest results of the 1.5 models I've tested.
Does anyone know where I can find a tutorial on how to write custom postprocessing scripts correctly in WebUI?The wiki is meh and I couldn't find anything useful. Even chatgpt seems incompetent.
Morning
>8 users generating 66724 images of "man in gym">7369 generations of the same prompt from 1 userlolneat
>>101580341This can't be true. Unless...
>>101580797This is so nasty and magnificent at the same time.
>join the doxxcord guise
>>101580817my stuff tends to have that effect on people>>101580865you can and should come back, I am willing to let bygones be bygones
not interested in joining the discord, i tend to spam too much, and go on about my depressing life, and i dont like that catjack person much
>>101580899community is good for your mental health
>>101580899Yeah, same for me.
>>101580899>i tend to spam too much, and go on about my depressing lifeIf you are aware of the problem you have no excuse to continue behaviors.
>>101580899Sorry you got triggered for being asked a normal question regarding your pedophile behaviorPlease stay locked in doors and away from children
>>101578212Is this supposed to be good looking?
>>101577617This is a conspiracy to build a moat. Because otherwise open source models will can just clone the outputs of Dalle3 and Ideogram and others and destroy their businesses by providing them for free.
>>101579014>>101578670holy jeebus
>>101580976>but enough about trani
so obviously i wont be rejoining.
the discord trannies seem really desperate for new drama lately. I wonder why
are there local installs of video creation?what kind of card could even process that, or would i need like 6 cards to even make it workdoubt my 2060 could deal with it
>>10158118812gb is the recommended amount of vram to have to do videos comfortably. 8gb is minimum for stuff like animate diff or svd. the newer video models are real piggus like tooncrafter at half precision taking 16gb
>>101581314man you really didnt make any progress at all in the last year besides using a tool someone else released lmaowhat a loser
>>101581314>the newer video models are real piggus like tooncrafter at half precision taking 16gbis that an inevitable trend going forward or is there room for optimization to keep memory demands lower?
>julien
schizoanon seems extra pissy today, must've had a hard week
waiting on top 1m subset to finish zzz
>>101581388>says the guy genning 1girl using someone else's ui
>>101581491please delete this
>>101581491kinda cursed but cool to see someone trying this out. is this liveportrait?
>>101581393NTA, but my guess is that VRAM will only be optimized for older models - there will always be a new generation that will create much more coherent videos and that everyone will want to use as they make the old generation look absolutely terrible in comparison.So nobody will want to use the optimized models; everyone will want to use the shiny new thing and produce things that look good.And that will always require a ton of VRAM. And Nvidia has no reason to make that affordable.
>>101581543yeh LivePortrait
>>101581393>is that an inevitable trend going forward or is there room for optimization to keep memory demands lower?that's really the goal but there are so few teams trying to do anything actually new. there was a runway dataset and training code leak apparently but I'd wait for a post about it for the news
>>101581543Is >>101577651true?
>>101581393>>101581650actually, here is the PC gaymer article https://www.pcgamer.com/software/ai/a-leaked-document-indicates-runways-gen-3-ai-video-generation-tool-may-have-been-trained-on-youtube-videos-and-copyrighted-content-without-permission/
cant wait until trani gets fired (yet again) :]
>>101575923>some people pretend that I somehow knew the node would have malware in itDo you have links to back up what other anons were saying? Or did you just make this up? Anon called you an idiot, not that you knew. And the fact that you doubled down and blamed everyone else is quite telling.
>>101581845>schizo desperate for drama
>>101581887>schizoElaborate?
>>101581023>Because otherwise open source models will can just clone the outputs of Dalle3 and Ideogram and others and destroy their businesses by providing them for free.The paper says the opposite, it literally says that if you spam synthetic data to train your model it will become absolute shit
>>101581773me too unironically
>>101581687is which part true?did I think synthetic training data is retarded?no. there are drawback but also valid applications of synthetic training datadid I ask for proof that it was retarded?maybe. if there is interesting studies or projects that provide insights, I'm sure I'd be interesteddid I remember asking for proof?apparently not. though its likely I would have came across relevant studies before anyone elsenow, on the topic of the linked article specifically, it seems people didn't even read the abstract. its not generally talking about synthetic data, its talking about recursively generated data -- particularly, long-term learning. it's also focusing on LLMs. you can assume there will be similar collapse problems with diffusion models but it wasn't explicitly mentioned or studied in this paperif you're interested in reading more on synthetic data in training, here's some additional papers to check outIs Synthetic Data all We Need? Benchmarking the Robustness of Models Trained with Synthetic Imageshttps://synbenchmark.github.io/SynCloneBenchmark/The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Betterhttps://arxiv.org/abs/2406.05184SynthCLIP: Are We Ready for a Fully Synthetic CLIP Training?https://arxiv.org/abs/2402.01832Scaling Laws of Synthetic Images for Model Training ... for Nowhttps://arxiv.org/abs/2312.04567Training on Synthetic Data Beats Real Data in Multimodal Relation Extractionhttps://arxiv.org/abs/2312.03025On the Limitation of Diffusion Models for Synthesizing Training Datasetshttps://arxiv.org/abs/2311.13090as you can see, its a very well-traveled topic in research with a variety of different approaches, applications, and conclusions. imo, the jury is still out
>>101581965tldr
>>101581965>he thinks his links non peer-reviewed papers is equivalent to a paper posted in Naturedamn, why are you always this retarded?
>>101581965Do you have any links to back up your claims? >>101581845 >>101575923
>>101581965>its not generally talking about synthetic data, its talking about recursively generated data -- particularly, long-term learning.Why are you moving the goalpost? This isn't the topic we're discussing about
>>101581965Ignore the schizo desu
>>101581965thank you for providing links and doing legwork, debo, regardless if I agree since effort should always be mentioned
>>101581965>did I ask for proof>maybe. >did I remember asking for proof?>apparently not.Wtf?
>>101582016forgot your gen >>101582021Who are you? I've never seen your images before.
>>101582033Nobody has ever seen your images because you've never generated any
>>101581965>there are drawback but also valid applications of synthetic training dataWhat do synthetic data offer that real images don't? That's the question we should be asking ourselves.
>>101581965>switching topics when losing an argumentlike clockwork
thread belongs in b with all the self dox and pedophilia innit
>>101581965>there are drawback but also valid applications of synthetic training dataExcept that using real pictures have 0 drawback, that's why it will always be a better choice than synthetic data, that's the point.
>>101582016I ignore him most of the time but that offered an opportunity for me to touch on that nature article and talk a bit about synthetic data>>101582021ty. the topic is interesting.also, idk if there's anything to agree/disagree with cuz the research still seems inconclusive on the worth of synthetic data>>101582056>What do synthetic data offer that real images don't?it offers a large wealth of data and usually more metadata with, since you know the parameters going into the outputs.its also an inevitable problem to contend with as more and more synthetic data seeps into datasets
>your opinion doesnt matter if i dont know who you are!!1
nyet comrade all image is same!!
>>101582088>it offers a large wealth of datathat's vague, can you elaborate on that?>and usually more metadata withwhat does that mean?>since you know the parameters going into the outputs.what are you talking about???
>>101581920What the Anon you're replying to is saying is that this paper is a lie, made up by the big AI companies to dissuade people from using their AIs output to train their own AIs to catch up.So they're saying that the AI companies made this paper and its conclusion up to stop others from catching up.No, I don't believe Anon is right.
>>101582088So you were lying in >>101575923 ? If what you say is true then it should be trivial to pull up links. Why do you lie like that? Do you think anon forgets easily?
>>101582017who's that?
>>101582124>what is RLHF
>>101582088>the research still seems inconclusive on the worth of synthetic dataonly peer-reviewed paper matter anon, and the only one we have on that topic is the Nature paper one, and that one says it's retarded to train models with synthetic data, and that's kind of obvious, why would you train a model with pictures that inherently have mistakes on anatomy, perspective and lightning when you can use high quality photos that accurately represent the world and reality.
>>101582140>anonDon't call him that. He is Debo, he's not an anon.
>>101582135>>101582124desu I would be ok to use synthetic data if you label those pictures as such on the training dataset. Like saying "this is a AI picture from Midjourney" or "this is an AI picture from Dalle3", that way you could make your model understand that AI pictures are different to real one and if you want something real, you could put "AI picture" on the negative prompt
>talking to thread schizo and expecting honest answersnewfags, all of you
https://reddit.com/r/StableDiffusion/comments/1ecjpvw/combining_sd_animatediff_tooncrafter_viggle_and/That's really impressive, probably one of the first short films that actually looks like it could be professionally produced.
you notice how peaceful and active the thread was when debo was asleep kek
>>101582140>when you can use high quality photos that accurately represent the world and realityWhich dataset is this? LAION? DataComp?
>>101582349I'm talking aesthetically, for the caption yeah you have to work and use something like GPT4V, that means not being a lazy ass and just scrap AI pictures that have anything inside yeah
>>101582367>synthetic captionsRetard
>>101582427You're praising synthetic pictures that has innacurate captions (because no AI generator will perfectly stick to the prompt you asked for) and aesthetic (bad anatomy, perspective and lightning) but you draw a line on synthetic captions? How can someone be such a retarded nigger?
>example #527472 why the basedbin of frenship should replace the discord in OP
>>101582499>>101582499>>101582499
>>101582033Do you like her? Let's call her Eva
>>101582304still too uncanny imho and the tracking was bad but that is on the editor
>>101581773I doubt it actually has ever had a job.>>101581965Ur gay and nothing you say or do adds value to anything.
>>101574985>>101574948NEW RELEASEX-men Evolution PONY LORAhttps://civitai.com/models/578582?modelVersionId=674346Best Quality BY FAR!!!!
>>101583970Tried remixing an image on my other lora showcase reel here:https://civitai.com/models/226324/rankinbass-studios-lorawith the Xevolution LORA.
>>101581314fuck's sake julien your gens are still the same fucking shit after 1 year of developments in the field?
END!
