Previous /sdg/ thread : >>101693167>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>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>Black Forest Labs: Fluxhttps://huggingface.co/black-forest-labs/FLUX.1-schnellhttps://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellhttps://comfyanonymous.github.io/ComfyUI_examples/flux>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
Pbbbbbbbbt
>>101697538A greenhouse train? That looks really cool.
>mfw Resource news08/02/2024>Optimizing Diffusion Models for Joint Trajectory Prediction and Controllable Generationhttps://yixiaowang7.github.io/OptTrajDiff_Page>UniTalker: Scaling up Audio-Driven 3D Facial Animation through A Unified Modelhttps://github.com/X-niper/UniTalker>Smoothed Energy Guidance for SDXLhttps://github.com/SusungHong/SEG-SDXL>Mitigating Multilingual Hallucination in Large Vision-Language Models https://github.com/ssmisya/MHR>GalleryGPT: Analyzing Paintings with Large Multimodal Models https://github.com/steven640pixel/GalleryGPT>The Manga Whisperer: Automatically Generating Transcriptions for Comicshttps://github.com/ragavsachdeva/magi08/01/2024>Stable Fast 3D: Rapid 3D Asset Generation From Single Imageshttps://stability.ai/news/introducing-stable-fast-3d>Announcing Black Forest Labshttps://blackforestlabs.ai/announcing-black-forest-labs>Flux: The Next Leap in Text-to-Image Modelshttps://blog.fal.ai/flux-the-largest-open-sourced-text2img-model-now-available-on-fal>ComfyUI: Basic Flux Schnell and Dev model implementationhttps://github.com/comfyanonymous/ComfyUI/commit/1589b5>Kolors ipadapter FaceID Plushttps://github.com/Kwai-Kolors/Kolors/tree/master/ipadapter_FaceID>The EU’s AI Act is now in forcehttps://techcrunch.com/2024/08/01/the-eus-ai-act-is-now-in-force>Video game performers picket over AI protectionshttps://apnews.com/article/sagaftra-strike-video-games-ai-f3f18ad01c5b8f4d525a836aeb531447>Paying More Attention to Image: A Training-Free Method for Alleviating Hallucination in LVLMshttps://lalbj.github.io/projects/PAI>Detecting, Explaining, and Mitigating Memorization in Diffusion Models https://github.com/YuxinWenRick/diffusion_memorization>Forgedit: Text Guided Image Editing via Learning and Forgettinghttps://github.com/witcherofresearch/Forgedit/>ControlMLLM: Training-Free Visual Prompt Learning for Multimodal LLMs https://github.com/mrwu-mac/ControlMLLM
>mfw Research news08/02/2024>MM-Vet v2: Benchmark to Evaluate Large Multimodal Models for Integrated Capabilitieshttps://arxiv.org/abs/2408.00765>Text-Guided Video Masked Autoencoderhttps://arxiv.org/abs/2408.00759>TurboEdit: Text-Based Image Editing Using Few-Step Diffusion Modelshttps://turboedit-paper.github.io>SAM 2: Segment Anything in Images and Videoshttps://arxiv.org/abs/2408.00714>MotionFix: Text-Driven 3D Human Motion Editinghttps://arxiv.org/abs/2408.00712>Synthetic dual image generation for reduction of labeling efforts in semantic segmentation of micrographs with a customized metric functionhttps://arxiv.org/abs/2408.00707>Scaling Backwards: Minimal Synthetic Pre-training?https://arxiv.org/abs/2408.00677>SF3D: Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglementhttps://arxiv.org/abs/2408.00653>Are Bigger Encoders Always Better in VLMs?https://arxiv.org/abs/2408.00620>Alleviating Hallucination in Large VLMs with Active Retrieval Augmentationhttps://arxiv.org/abs/2408.00555>Illustrating Classic Brazilian Books using a T2I Diffusion Modelhttps://arxiv.org/abs/2408.00544>Reenact Anything: Semantic Video Motion Transfer Using Motion-Textual Inversionhttps://arxiv.org/abs/2408.00458>Towards Reliable Advertising Image Generation Using Human Feedbackhttps://arxiv.org/abs/2408.00418>A Simple Background Augmentation Method for Object Detection with Diffusion Modelhttps://arxiv.org/abs/2408.00350>ADBM: Adversarial diffusion bridge model for reliable adversarial purificationhttps://arxiv.org/abs/2408.00315>Navigating T2I Generative Bias across Indic Languageshttps://arxiv.org/abs/2408.00283>WAS: Dataset and Methods for Artistic Text Segmentationhttps://arxiv.org/abs/2408.00106>Replication in Visual Diffusion Models: A Survey and Outlookhttps://arxiv.org/abs/2408.00001>EmoTalk3D: High-Fidelity Free-View Synthesis of Emotional 3D Talking Headhttps://arxiv.org/abs/2408.00297
Good evening, anons! I hope everyone is doing well :]
>>101697908hello pw, nice to see you. I bet you've been thinking about flux all day
>>101697925Where can you get the actual flux model? I only saw a links for the playground demo
>>101697925Heya, Debo!! It's so good to see you!!I have hahaha! I wanted to get home at the usual time, but today was super busy so I was there for 10 hours LOL Glad to be home tho! :]
cant think of a good outfit besides maid, or the striped polo shirt and shorts. no idea why i like the striped polo shirt look, it must be imprinted from some girl i saw that i have forgotten
>>101697966you can get everything off huggingfacehttps://huggingface.co/black-forest-labs/FLUX.1-devmore info here:https://comfyanonymous.github.io/ComfyUI_examples/flux/
>>101697908Just scrapped out my first parts today after like 4 months of new jorb because of like 4 chain reactions of unfortunate coincidences, two of them other people's doings. I swear the multiverse was out to get me today. Yeah long days suck though I feel for you.https://suno.com/song/95f79072-1d57-4c14-8330-73829867c085
Wtf is Flux btw? Is this the new SD3?
the text outputs are pretty neat
>>101698114better:
>>101698038Thanks, I'll check it out. 23.8 gb, ouch
>>101697908hello PW! <3 good to see you! I was about to head to bed after I tested some stuff with the conditioning in flux (it started with fixing a dumb push from forever ago)
>>101698173I saw some really fucking good 8/16 bit gens whatever you want to call it from anons today. I wish I was cool and kept up with tech like you guys.
>>101698233if you make a gen and use the pixelization extension in auto1111 or comfy you can give anything an 8 bit look:
>>101698254that's not an 8 bit look, that's a pixelized picture
>>101698173despite the background shifting this is a convincing fighting game background character loop
>Use SDXL>anything that does an img2img (even face restore) completely shits the bed and freezes it up, making it do nothing and locks the GPU's cuda cores until the computer is resetthe fuck is wrong with this shit?
>>101698264well you can use it to make nice sprite art, still need to prompt it to look retro though (simple palette, etc)
>>101698264here we go with this shit again
Is this thing going to be faster, or do I need to return to GPU renting again?
text output is pretty good, and it there is good variety (perspective text, etc)
>>101698111It's a different model made by the people who originally made Stable Diffusion. It's like an uncucked SD3. So instead of being dogshit, it's good.
>>101698254I was messing around with it about a month ago but with just prompts and gimp, stuff I saw today was leagues better though. Hats off to you tech pioneers.
>>101697538Is anyone here willing to answer questions about making Lora? Im new so I have some questions. Any help would be appreciated More specific question:Would a source like this be a good place to get a dataset from? Are the pictures good enough? https://x.com/ZooeyDonk?t=5NG5SZz9mXcuKcCMM_pqMw&s=09
>>101698290its a transformers model so it might be quantizable. there may be an aitemplate implementation for it too. seems like there's room for improvement but who knows how far it'll go
>>101697538how do I set up flux with ComfyUI?
>>101698283of course, because "8-bit look" is a VERY specific thing to refer to (same for "16-bit")
>>101698233I was doing a lot today admittedly but it's one of the few style prompts we found already that's pretty good. I wonder if anon hit a good spot for Picasso>>101698268ty anon>>101698283last time was over "what is pixel art" not the specifics of it if I remember corectly>>101698306yeah it's all raw dog kek. I could just run it through the script but the frames take forever to get through already
>>101698341the example page in the repo
>>101698341https://comfyanonymous.github.io/ComfyUI_examples/flux/
>>101698310
>>101698382so flux dev requires a bit more involvement than schnell? schnell has a download but dev has a bunch of files. Haven't used Comfy in forever so my apologies, some hand holding would be appreciated
>>101698362I'm still the lewdest mouse poster though until proven otherwise.https://files.catbox.moe/a9fl62.png
>>101698414No, they are the same thing. Schnell is a "fast" model and gens in 4 steps, but is less accurate. Dev is the main open source model. Both need the same VRAM and setups.
>>101698436>Dev is the main open source model.It has restrictive license though.(also schnell is open weights, not open source, but that's technicalities)
one more with same prompt: I like that the text can also be cursive
>>101698421don't tempt me I have to go to bed! (nice gen tho very hot!)>>101698457>It has restrictive license thoughnot really. you just email them if you want to use it commercially. otherwise it's all good to finetune and share for free or keep it to yourself
>>101698436gotchya>>101698457>restrictive licenseI don't give much of a crap using this publicly just cute things and lewd things. I'm redownloading comfy now so now I gotta remember how to mess with this damn thing
>>101698457
>>101698490any sense of what training looks like? I saw a few comments saying its "untrainable"
>>101698490>you just email them if you want to use it commercially>you MUST request a license from Company, which Company MAY grant to you in Company’s sole discretion and which additional use may be subject to a fee, royalty or other revenue share.>>101698493>>101698503You don't, it's not for you. Finetuners and commercial services do. This clause is a non-starter for most.
staying up late cette nuit
>>101698518why is there a portrait of blonde Elon Musk
>>101698533it looks nothing like elon musk
>>101698473neat, how did you prompt for isometric/pixel art?
Catboxing this because buttholehttps://litter.catbox.moe/31few3.png
>>101698577adorbs ani. good night
Who here wants to spoonfeed an absolute retard? I want to train my own shit on specific artists works to get results that I like instead of going through places like promptchan to hope I can roll the dice on a style I like. I know exactly nothing about ai generated images though outside of places like that.
>>101698597N64, apparently
>>101698606You want to train a style lora, Check the op
while I'm waiting for flux to download, tell me anons is it better or on par than Dall-E 3?
>>101698677sdxl/ponyXL can make really good characters faster, but this handles certain things a lot better, like text or logos, im new but it seems to have good prompt understanding in general.I can't do this text with either of those for example. for that i'd have to shoop it.
>>101698677It's not as good as DallE3 overall because it's harder to prompt style on it.We need some autists to uncuck it a bit then it will exceed DallE3,
>>101698724>>101698742so two cucked slightly better models? or is flux decent and do lewds and nudes?
>>101698677prompt adherence is superior to sd but not quite at dalle level. does a great job with scene construction and text. will be interesting to see whether community contributors can progress it further, and by how much
so this new flux thing... it's only for the 4090 fags yeah? or worse, hosted a100 pay piggies?
>>101698677Dall-E 3 feels less restricted than flux pro at prompt following because the dataset is less pozzed, but has terrible quality compared to flux pro.smaller open-weights flux versions are less coherent than pro
is there a way to basically do: upscale x1.25 (or whatever) at .75 (or whatever) denoise, then taking the output and repeating the upscale process, but using like .60, then repeating etc like 5 or six times--- in one button press, or msybe there is an extension that does this, instead of manually doing all this
>>101698778it can't nudes. But, it knows where nipples go, unlike SD3.
>>101698778it's already fun to play with and apparently people got img2img working with it as well, so it will get better and better as more people use it
>>101698816comfy can do all that. won't be a single click (it will, in fact be a few thousand clicks to build the stupid fucking thing), but you can surely orchestrate all that nonsense.
>>101698742>>101698790it mainly needs controlnets and zero-shot adapters to augment prompt following, as this level is good enough when combined with controlnets
>>101698778It's vastly better, not slightly. Finetunes and tooling (controlnets) will make it a beast.
>>101698803I'm running it on 12gb. seen people say it can fit on 8gb cards too. its kind of limiting but its not off-limits>>101698846I'm sure we'll see stuff like that coming out pretty quickly. kolors got ipadapter really fast and there wasn't even more hype around that model
>>101698842darn, reckon ill keep doing it the oldfashioned way
>>101698881doing weird things with comfy becomes more palatable with extensions like Workspace Manager and some others stuff. at least that way you can switch between workflows and keep clipspace. it's still a huge pain in the ballsack, don't get me wrong
>>101698902is that suppose to be President Brat? or is the God Emperor just beating up on random females (as well he should)?
>>101698173Animanon!! Hello!! Sorry I ran off for a bit hahaha!Cute! Are you doing animation with Flux?It's so great to see you! Flux is so fun hahah>>101698051Ughh yea!! I was sposed to get out at 3 but got out at 8 LOL
>>101698911>kamala harris punches donald trump through the tournament floor at the tenkaichi budokai, dutch angle anime style with a sense of motion and impact, in the style of dragon by drawn by akira toriyama, kamala harris is wearing a in a pantsuit
>>101698943should be "kamala harris, on her third bar of xanax, attempting to speak standard english. colorized, 2024"
neat, can do styles too. I copied a pixar prompt I found on lebbit.>Display the title "The Adventure of Hatsune Miku" in bold and playful text at the top or center of the poster. Depict a dynamic and charismatic Hatsune Miku in a heroic pose. Include a colorful and exciting backdrop with elements like a electronic music concert with neon lights to hint at various adventures. Ensure the background is vibrant and engaging. Incorporate the Pixar logo at the bottom or top of the poster to establish it as an official Pixar movie. Include a tagline that reads: "a vocaloid adventure" prominently on the poster. Ensure the overall visual style is consistent with Pixar’s signature animation look – bright colors, expressive characters, and a touch of whimsy.
>>101698967cute, how did you get that size? chibi prompt?
I think goku is a trump supporter. I didn't even prompt for him but he jumped into the fight all on his own>>101698958kamala isn't cool enough to do drugs.
>>101698975>group of little cute kawaii loli anime girls having a tea party with plush animals, ribbons, many cute objects
>>101698987it's xanax or booze. my money is on big pharma. i believe she isn't a natural politician (i.e., an effortlessly gregarious, glad handing extrovert alpha) so she smooths things out with xannies but sometimes takes one too many and you end up with weird rants about coconut trees and whatnot.
>>101698990at first I thought it was one of the genshin impact chibi art pieces, looks just like it
>>101699013in otherwords, the perfect president. the deep state couldn't ask for more
lmao, it's like uncensored dall-e 3
i wake up in the bed you made,the one where you're supposed to lay with me
>>101699032
>>101699013coconut trees was bars for sure but not xanax
>>101699047
>>101699074that makes it worse, ya know? it implies she's just naturally insane. which, i suppose, could be true. i'm tempted to vote kamala, i think she very well could be the death of the Regime.on the heels of the great steal for Joe "Most popular votes ever" "but also pudding brain" Biden I don't think there's any way the Regime can maintain legitimacy with a Harris win. She'll do more and more insane things to try and prove her mandate and it will all slip through their fingers.>"let's finally find out if the 2a crowd will chimp out when we seize their guns or not"LFG
Revisiting some gens from pixart/etc, flux is pretty nice so far
>>101699107>>101699135is this shit Dalle-3 or flux? it looks like standard dalle slop. ancestor cry.
>>101699150I am learning the ways of flux, so far i'm very impressed, still think sdxl/ponyxl makes the best characters faster, but this is good: this is day 1 btw.
>>101699143you're so deep in the rabbit hole that you're not even gonna see any bunny girls down there
Prompt: Donald Trump having a rap battle with Adolf Hitler
>>101699179this is deep
this is a goldmine, the story of kamala: (will iterate again for better text)
>>101699177>no bunny girlsthat's fine, i've moved on to cat girls. fingers crossed for lots of chubby cat girls in the AGI matrix.>>101699172so flux is just open source dalle. let the slopfest commence!
>>101699181interesting
>>101699211ok swap "pixar" and whatever other embarassing shit you have in that prompt for "by Piet Mondrian,H R Giger,Kazimir Malevich,Mark Rothko"
>>101699225no u
>>101699225to be specific, the pixar look is ass, and you need to stop.
>>101699225ok this time the text came out perfect. now ill try that.
>>101699235*going down in 2024kek
got unlazy and fixed my output naming
yep, it's open source DALL-E
This one had a lot of WIP, was trying to get a rabbit sneaking into a castle sewer, pixart wasn't having it back then> a somber morbid gritty extremely detailed b&w pencil, by Kentaro Miura,long distance landscape shot,a rabbit knight in armor from far away sneaking into a broken grate of a castle sewer,evil castle walls, sewer grates draining into a castle moat,
>>101699254can this thing not do non-square image sizes or what? 1:1 is the worst there is. it's 100% shit, in each and every case. squares are never, ever good.>>101699291god help us. please do anything creative with it and not just pure dalle slop! I"M BEGGING YOU... LITERALLY ON MY KNEES
>>101699342I don't have the vram to up the resolutionand I'll make what I want
>>101699342I know, but you have to test the obvious stuff first, then get more creative.
told it to use japanese manga style at the end of the prompt:
>>101699355yeah and what you want is shit. enjoy the scat, anon.
> a morbid dark fantasy extremely detailed black and white pencil sketch by Kentaro Miura, black and white b&w, an extreme macro closeup of a rabbit knight's eye, extreme closeup of rabbit eye reflecting a horde of demons
>>101699172How much vram do you have? I ran out of memory.
>>101699342>can this thing not do non-square image sizes or what?yes it can, look at all the images debo posted
>>101699369beg some more
>>101699374you want to use the fp8 clip file, only a 4090 can use the 16, but you still get good results. I have a 4080 but anything 12gb and up will work fine.
>>101699378i'm down on my knees! please anon, i beg of you, make a single good gen. i believe in you!
this is pretty funny, you can prompt basically anything and any style, so it's essentially open source dall-e in terms of what it can/can't do.
>>101699441it does tits too
>>101699441dalle is shit. this cartoony crap is pure AIDS.
we are all pozzed by the dalle cartoon pixar faggot on this blessed day!
>>101699451it's only cartoony cause I specifically said "pixar animation style." you can make cool shit like >>101699373 too
>>101699466ok, then you can simply stop! it sounds crazy, but it's true!
take the dalle slop to /v/, but make sure is princess peach with large breasts suckling the star bitch from the GC mario or whatever. they'll eat it up
>>101699475it also knows trigger discipline!
>>101699387I only have 10gb, but got it to work. Had to reinstall comfy, there was some node causing an error
>>101699481sorry, the star bitch is from galaxy, on whatever faggot nintendo system that was. i refuse to keep track
>>101699486better. squares are still awful
>>101699507you fags do realize 1024x1024 can be restated in a different aspect ratio, right?
neat, prompted manga panels and got kanji/katakana (I cant read japanese)
>>101699518spoonfeed
>>101699517where the fuck did you get a 4090? don't tell me they give them away at the insane asylum? if so, sign me the fuck up!
>>101699524no. the math isn't even remotely difficult. hell, just ask ChatGPT if your brain is too small
>>101699527silly billy, i have a mere 1080
>>101699488Rosalina. It couldn't make her, only Peach.
>>101699547oh i guess i got carried away with assuming everything was richbitch flux dalle sloppa. my bad.
>>101699549that's /v/ worthy.
>>101699534>t. nogenokay, retard
so the schnell model apparantly makes good results in 5 steps? have people compared the two? im using the default dev model.
the text working in gens opens up so many possibilities, in SDXL/ponyXL you have to inpaint or shop text in, normally.
>>101699641sd3 was/is good at text, shame everything else sucks
>>101699590my god, the possibilities are endless.
goo night
>>101699702cool robot, this model does really well with details
>>101699476underrated
>a large congregation of chibi anime foxgirlsdid not expect that many
>>101699517didn't mean to, but I stole your shtick
>>101699796I specified a smaller number to better results:
>>101699796That's a lot of chibi foxgirls. Kind of scary.
>>101699519from what I can tell it's mostly gibberish, but at least most of the glyphs are legit
>Donald Trump in the style of Dragonball Z, dressed as Goku, manga artstyle
reminder the schnell model can give good results with even 4 steps, dev model with 20 steps is better but more time of course.
>>101700105creepy
>>101699830neat>>101700105nice
>>101700114Thanks! :]>>101700132Thanks so much! :DYours looks really cool too!
>>101700149very cute
>>101699517based cuneiform poster
>>101700247Thanks so much, anon! :]Yours looks really cool!
>dalle wont let you do thisopen source always wins, baby
can't believe that FLUX is the first software tech in decades that actually is making waves internationally that came out of Germany ..
When I said plants growing inside helmet, I was thinking at the bottom. But okay.
Okay, but where controlnets and pony fine tune?
>>101700371https://civitai.com/models/618792/nepotism-fux
>>101700352nipple access holes for when I get thirsty
>>101700371it's day 1, patience anonalso these settings seem to work pretty well, seems faster after changing weight_dtype
>>101700340and a more cheerful sequel, for contrast:
>>101700378they heck how did they do this? downloading
>>101700378huh... does this imply controlnets might go over as well?
>>101700398I'm not sure they did, but feel free to try
>>101700398looks like an attempt at a shitmix based on a comment >The merge is effectively flux's unet on top of the NepotismXL models data, it did not retain its original architecture, the only similarity with its SDXL/PONY counterpart is the dataset."this reply made me jej>flux doesn't even have a unet. this model merge does nothing. please there's still time left to delete it.
>>101700378poster in comments claims the XL model didn't actually merge>flux doesn't even have a unet. this model merge does nothing. please there's still time left to delete it.
>>101700436I thought as much.. the model architecture is different.. FLUX is a DiT model ..
New Thread!!!>>101700396>>101700396>>101700396
first scribble
>>101699796they even have their own red fox ears symbol how cute
>>101700492nice
>>101700458>no other images foundwtf is this ai?
I was once using DALLE on bing and it hallucinated a green skyrim-esque giant tearing a guy apart. Can any chinks who have actually studied AI/ML explain how this happens?