Previous /sdg/ thread : >>101700396>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
float8 weightsnicethis is just the FluxSingleTransformerBlock but FluxTransformerBlock will be about the same speedso it's already twice as fast as before, smaller weights to load with each module makes that much differenceworkspace changes will help moreinb4 hateful people moaningnew models always bring them out
>>101708609It's hilarious hubris honestly, everything can be reversed. The question is how difficult it will be to get the compute but honestly when you have a very obvious end goal and target it makes research easy because the research simply is>we want to train Schnell on a 24 GB GPUMany people will be taking a crack at it and we already know the power of boners and there are many PHDs who literally have access to H100s provided by their universities.As far as I know, to reverse Schnell all you need to do is be able to fully fine tune the weights. This is a target that can be achieved because after all, they had to be able to train it themselves.Watch as someone figures out how to shard it into bite sized 8 GB pieces.
>>101708688link?
>mfw Resource news08/03/2024>TryOnDiffusion: A Tale of Two UNetshttps://github.com/fashn-AI/tryondiffusion>Nvidia reportedly delays its next AI chip due to a design flawhttps://www.theverge.com/2024/8/3/24212518>ComfyUI Frontend Modernization: Transitioning to a New Era on August 15, 2024https://github.com/comfyanonymous/ComfyUI/issues/4169>CEO of Invoke says Flux fine tunes are not going to happenhttps://www.reddit.com/r/StableDiffusion/comments/1eiuxps>ComfyUI-FLUX-fal-APIhttps://github.com/gokayfem/ComfyUI-FLUX-fal-API08/02/2024>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 T2I Modelshttps://blog.fal.ai/flux-the-largest-open-sourced-text2img-model-now-available-on-fal>ComfyUI: Basic Flux Schnell and Dev 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
>mfw Research news08/03/2024>Image Super-Resolution with Taylor Expansion Approximation and Large Field Receptionhttps://arxiv.org/abs/2408.00470>Localized Gaussian Splatting Editing with Contextual Awarenesshttps://arxiv.org/abs/2408.00083>Hierarchical Conditioning of Diffusion Models Using Tree-of-Life for Studying Species Evolutionhttps://arxiv.org/abs/2408.00160>SynthVLM: High-Efficiency and High-Quality Synthetic Data for Vision Language Modelshttps://arxiv.org/abs/2407.20756>Real Face Video Animation Platformhttps://arxiv.org/abs/2407.18955>ObjectCarver: Semi-automatic segmentation, reconstruction and separation of 3D objectshttps://arxiv.org/abs/2407.19108>Multi-Expert Adaptive Selection: Task-Balancing for All-in-One Image Restorationhttps://arxiv.org/abs/2407.19139>Exploring the Adversarial Robustness of CLIP for AI-generated Image Detectionhttps://arxiv.org/abs/2407.19553>Advancing Prompt Learning through an External Layerhttps://arxiv.org/abs/2407.19674>VolDoGer: LLM-assisted Datasets for Domain Generalization in Vision-Language Taskshttps://arxiv.org/abs/2407.19795>Mixture of Nested Experts: Adaptive Processing of Visual Tokenshttps://arxiv.org/abs/2407.19985>Perm: A Parametric Representation for Multi-Style 3D Hair Modelinghttps://cs.yale.edu/homes/che/projects/perm/>ML-Mamba: Efficient Multi-Modal Large Language Model Utilizing Mamba-2https://arxiv.org/abs/2407.19832>Adversarial Robustness in RGB-Skeleton Action Recognition: Leveraging Attention Modality Reweighterhttps://arxiv.org/abs/2407.19981>Exploring Robust Face-Voice Matching in Multilingual Environmentshttps://arxiv.org/abs/2407.19875>MaskInversion: Localized Embeddings via Optimization of Explainability Mapshttps://walidbousselham.com/MaskInversion/>Task-Adapter: Task-specific Adaptation of Image Models for Few-shot Action Recognitionhttps://arxiv.org/abs/2408.00249
>>101708688oh is this AIT, you're compiling each layer to it's own program and then you can run them sequentially? is that how that works?that's great i was just smoking weed and thinking about how you could probably ring buffer the weights and swap em out of gpu as they get processed to save on vram
>>101708758when its ready, its for developers only atm>>101708771yes, as a workaround for the model being so big ive split it, each layer runs separately. atm im binding the constants to the module but i will also test it using the set_constants function which would then only have 1 module per block type
>>101708744You do realize when SD came out there was zero tools to train it right? You realize textual inversion did not exist when SD came out, right? Fucking tourist.
flux sampler/scheduler comparison grid when
>>101708771>i was just smoking weed and thinkingthats cool. when I smoke weed, I watch cartoons
Installed flux last night and generated a single 1026x1026 image and all it outputed after 10 minutes was a blank black image. I'm guessing 1070ti isn't enough to use this model, yes Im poor.
>>101708830>You realize textual inversion did not exist when SD came outit did though
>>101708930Then you are a bigger retard than I thought. You can shut up now. No one asked someone with 90 IQ what they think could happen.
>>101708935https://github.com/rinongal/textual_inversioncheck for yourself
I feel like we deserve flux after the sd3 disaster.
>>101708906Wikipedia, Blender, Mozilla, Linux, Python and so many others are nonprofits.I believe that truly revolutionary local AI shit will come when we get a nonprofit org that gets funding all over the world to deliver something good.It's what SAI tried to become, but failed miserably.
>>101708971Damn I wish I could use it, whats your pc specs?
best site for doing videos?
>>101708943Gee anon I wonder which image model they're customizing. Did they make this code before Stable Diffusion existed? Are you stupid? Are you a tourist? Help me understand.
I trying out the NAI v3 gen on sd-webui with the extension, anyone knows why I'm getting washed out colors all the time? Sorta like the ones without VAE but I'm not sure about setting that up with it
>>1017090024080/7800x3d/32gb ddr5, but this model will run on any 10-12gb+ cardthere are also online options too I think
>>101709022luma
this one turned out better
>>101709024any latent diffusion model. textual inversion was released before sd was released. ironically you're the tourist, don't try to pretend you were around at release.
>>101709072that's not very safe
Could make some good money off tshirts this season
>>101709088Actually I was, I ran SD from the command line. That's how I know you're a fucking moron because you think we were training customized models day one. Fucking retarded stolen valor zoomer probably was in middle school when SD came out.
its pain to fix pose imagetakes 5 - 10 minutes
>>101709103my first embed was released 25th aug 2022 though, not quite day 1 but close enough
>>101709144then you would know it was leakedretard
gotta stop 1girling
First time comfy user.I load the flux workflow, choose the model files, hit queue prompt, the entire thing crashes no error logged.Terminal just says pause.What the fuck am I supposed to do?I have 4090
I'm generating scenes with +2 people and using pic related to try to detect the people. It's working good and properly detects each person with a proper silhouette. The issue I'm having is that they come out as one entire segment rather than separate segments. Is there anything that can be done to either separate the people into individual segments or just have the detector work on a specified region of the image?
https://litter.catbox.moe/640z6x.png
>>101709160whats that got to do with anything
>>101709252You're saying TI was there day one, which it wasn't, not even close. This in relation to a new SOTA model which just barely emerged from the womb which you're now doomposting about as if that's relevant to anything. My point is SD spurred a shit ton of research and things rapidly changed and you're in here probably as an employee of Flux acting like your model can't be touched.Your model will be raped and any hopes you have with Pro will be tossed when everyone realizes show bad faith you were all along.
>>101709192>the entire thing crashes no error logged.I'm not sure thats possible
the text working consistently well is what makes this model unique. SDXL/ponyXL can do good characters but not text like this (yet)
>>101709192>I have 4090You need at least a 5080 Ti to run FLUX. Cheap cards won't make it.
>>101709270all these schizo /pol/ tourists. its so tiresome
>>101709270I'm not sure leaning into schizo posting is gonna help you make your argument
>>101708830Textual Inversion and LoRAs existed before Stable Diffusion. They were designed for use in LLMs, not in latent diffusion. We were using TIs to "train" SD since it was leaked.
>>101709306I say this and look like this
>>101709311Prove it, show the TIs with a post number. I was there, it was just Emma posting.
>>101709310>>101709305wow discord faggotry in sdg noo, definitely no collusion hereanyways, enjoy flux boys, it's going to be trained
people are already finding ways to get it working on 8gb, just be patient anons
>>101709346>it's going to be trainedWhere and by who?
>>101709365No one was going to full fine tune SD, impossible they said.
>>101709355>people are already finding ways to get it working on 8gb, just be patient anonsAnd patient you will learn to be, because you will wait 5-10 minutes per image.
>>101709280It just dies
>>101709327Oh OK, I will go back in time and think to take a screenshot of a post because some day in the future a retard will be a faggot
>>101709378the schnell model can generate decent stuff in only 1-4 steps. but if you have the vram, dev works best (so far)
>>101709391can't wait to see a time stamp a week or more after the initial leak
>>101709391>because some day in the future a retard will be a faggot
>>101709391you're in luck anon, every post on /g/ is archivedthat means you can go and look at what was posted on August 19, 2022
can you explain to me this particular one? the winner of that match was just a normal woman with high t right, not actually trans
it is neat though how sd3 was a bit of a huge letdown (cause it was censored, thus the training was shit) and now we get this, which is basically mini dalle-3.open source is the way.
>>101709389https://openart.ai/workflows/maitruclam/comfyui-workflow-for-flux-simple/iuRdGnfzmTbOOzONIiVVI used this guide, also make sure the unet/clip files have the same filenames or use the arrows to select them if differentif it acts up, download the comfy portable .zip and do a fresh install of that, then move the files over.
>>101709451I hope all this AI shit gets axed the worst way possible
>>101709451>open source is the wayMarketing gimmick. They have zero interest in supporting the scene in creating the next base for porn finetunes. In fact, they are betting that it will never happen, gloating.Remains to be seen if autists can bruteforce and build on top of these models.
>>101709493the meme stuff is just to test functionality, it's capable of really neat stuff in terms of art styles. It can make manga panels and even do japanese characters. the text outputs are very good so this is a big step forward for open source models.
>>101709449yes
>>101709451The upside to Flux is that it's local, so normies get filtered and there won't be a mass flood of these types of images. A small number of quality ones is better than what happened with dall-e.
>>101709517that's not what i meanmark my words the few degenerate scumbags are going to ruin AI for everyone else but you already know that
>>101709525>The upside to Flux is that it's localNormies use Flux Pro, which you can't (locally).
>>101709525and normies can't figure out comfyUI even though you can just click and drag a jpg to copy a workflow.>>101709545nah, the opposite: pony models were meant for making hardcore porn, right? yet they can make amazing anime gens, even though you CAN make porn doesn't mean the model is ONLY good at porn.
is sai kill?
>>101709565this is also why sd3 fucked up, they censored the model, so it's shit at anatomy other than faces. if you dont teach an AI model how figures work, it can't make them properly.
>>101709493You can hope for all you want but it's not going away. Settle in and enjoy the ride.
>>101709565Yotsubest!
is there a comfy node that shows the image before it's complete, like every x number of steps it updates
>>101709640you can enable preview in the settings
>>101709545Pony already can do degenerate shit and we've survived.
>>101709098
It's over.
>>101709668good
>>101709652and it's a great model BECAUSE it can do degenerate shit. if you can render a body doing almost anything, you can do the same for fully clothed characters too. without proper anatomy training, gens don't look as good.
>>101709677ai gens blobs of color, it does not understand we're naked under clothes
>>101709652Pony sucks balls. Oversaturated colors. Fucked up hands and faces.
>>101709687
>>101709686even this font is just dots, like noise, that forms text. latent noise is the same, it forms an image. it knows what boobs are because it learns that from training data and concepts.
>>101709687use autismmix-confetti, it's much better for anime gens than the default model.
>>101709787no, when you say "clothes" it hallucinates blobs of colors associated with clothesnever once does it generate a naked person under those clothesno, it generates a plausible image of blobs of color associated with "clothes"
>>101709687base pony is shit but good for training loras and then using the loras on other bony based checkpoints there's so many style loras than im sure you'd find what you're looking for
>>101709819>blobs of colorsAs opposed to the blob of fat that is your mother?
>>101709819yeah, it doesn't know what naked is, that's a human concept. to the model it's just an array of dots.
>>101709686yes it doesn't "understand" that we're naked underneath out clothes but it doesn't need to, please don't be retardedhow do you think the weight dimensions happen? when training pictures of people will adjust weights that are related to people, pictures of naked people will adjust weights related to naked and people, clothed anatomy improves with nudes in the dataset
>>101709842it also doesn't know balls bounce
>>101709855That's cute but flux has no genitals or really any porn yet somehow manages which destroys your premise
>>101709819well, I have to disagree.When im genning clothed 1girls, and observing the preview, i'd see the outline of a naked girl's body, and then the dress get's genned over the naked silhouette.
>>101709211Impact has a 'SEGS to Mask List' node. Maybe try that.
lost, promptwise, gonna try to play nomanssky again
>>101709773Is your picture generated from Pony?
>>101709882prove it, you can't
>>101709903Autismmix, a Pony derivative.When people here say Pony they typically mean Pony and its derivatives, not just base Pony.
>>101709904im busy baking a lora rn so you're right i cant prove it (right now) :^)
>>101709893one of my favorites. I was just building a base at the center of the galaxy
now this is podracing.
>>101709889Wouldn't help, the issue itself is the SEGS not coming out as individual ones. It's always one that has all my characters.
Hopefully the black forest labs text to video can at least come close to Runway Gen 3. That would be the real game changer.
>>101709976Anon, she asked you to stop this >>101702717
>>101710079not a footfag but wowwww that's really good
>>101710102id complain about zoomer meme miku's too
>>101710114The caveat is that 90% of generations in runway Gen 3 devolve into unusable junk. Only 10% are really good, so it's still a lot of trial and error.
I'm worried about my cpu, it's been at 85 degrees for about 18 hours of genning now
>>101710233new paste
>>101709178I challenged this ultra hard pose
>>101710247i'm very tempted to get an AIO cooler. it's not like the £50 I'd spend on it comes anywhere close to the price of a 3090 which is a much better investment
weird cutlass bug when add_*_proj is cast from float8, idk, doesnt make much difference to keep those in float16anyway this is both FluxTransformerBlock and FluxSingleTransformerBlock, so yeah about twice as fast as it was earlierloading/unloading is still slowing it down so ill try the set_constant method next, but this method is extremely beneficial for ramlets as well as vramlets
>a centaur girl wearing one-shoulder clothing, carrying a long rifle, single bare shoulder
>>101710298>weird cutlass bug
Dig the shirt here
>>101710336nice
I really like the color combo here, but I've been struggling to replicate it since. The hard part is getting the right shade of brown
>>101709668Random pineapple opinion.
>>101710336Cool
>>101710452thats the ceo...
>>101710471well what the fuck does he know, he's a suit
>>101710485>>>/pol/
>>101710510that's right, nothing
>>101710429omg it migu
>>101710530das rite
Im currently testing out flex pro, and have an account i can spend some money on, pls ask for prompts. This here is:a giant billboard at night that reads "sex problems? not with Xera!
>>101709668>he's directly challenging the autism of the pony community
>>101710617>billboard alongside a small dirt road
>>101710654i think you dont understand how advanced models work, "at night" means that you want this in a setting where its clear that its night. How else would you logically show its night?
>>101710694What the FUCK does a dirt road have to do with night?
>>101710617>goku powerbombs hatsune miku through the tournament floor at the tenkaichi budokai
>>101710741roads only turn to dirt at night
Flux seems to be really good at specific styles
>dat fur on the forearms
Couch potato
>>101703851>>101707455I've implemented SEG for ComfyUI/Forge in https://github.com/pamparamm/sd-perturbed-attention, not sure if it's better than PAG tho
>>101710999Kek literally me
>>101710957nta, but thanks
>>101710957errm, its Smoothed Energy Guidance but you named the repo Perturbed-Attention Guidance
Prompt: 'faggot'
>>101711171It's the same thing basically. SEG is a variation of PAG
might defork it when i get to 100
>>101711200>>101711186Try to prompt the reflection in the glasses to something specific.
Is there a repository of information about SD techniques that are a bit more intermediate/advanced/obscure? For example, I recently learned about Differential Diffusion that basically converts any SD model into a model capable of inpainting. And now I see folks discussing something called Smoothed Energy Guidance, and I'm getting imposter syndrome kek.
>>101711244nah, too much stuff happens all the time for there to be any sort of central repository of information. most information is spread ad hoc through social media like twitter, reddit, discord, and here. all you can really do is search around and try to dig up conversations on things you want to learn more about
>>101711221
Depict a dynamic and charismatic Hatsune Miku dancing on a stage. The image is in the style of a comic book from the 1930s. Miku is saying "Hello, SDG!" in a white speech bubble.
>>101711535Hello Miku
>>101711535
>>101711574
>>101711535now make goku powerbomb her through the floor
>tfw flux doesn't know celebrities
>>101711607one more, but a diff style prompt
>>101711639yeah I was trying to get a missy elliot rap video. this is missy elliot
COME VISIT NEW YORK
Anyone else disappointed that Flux appeared so great but turned out to be dead on arrival because it cant be trained?
>>101711718I think it's not impossible to train. Just potentially very hard, especially considering the ungodly amount of compute power and VRAM required.
>>101711718sdg is waiting for ldg's apology
>>101711734people can rent super high end GPUs, just give it time
>>101711696>>101711708sexo
>>101711718Releasing the models to the public was a publicity stunt, yeah. They made a great model though. That must be admitted.
>>101711696>>101711708WTF how did her coat become transparent?
>>101711718I think personal loras are no-go, but we'll probably see finetunes eventuallyI can live with it
Nice, so we're back to using 1.5 and SDXL as our models again now that Flux can't be trained. It's over for us localchads.
>>101711734Even loras and controlnets would be enough. is it even possible?
any reason why my bake is slowing to a crawl?first time this is happening
>>101711777attempting a pony pass on a flux gen and it's just not comparable. Every time it makes it worse, even if it can attach some normal looking nipples. If Flux is trainable it's over for SD. >b-but some random faggot saidnobody cares
i miss schizo anon
>>101711813Right is original pony?
>>101711836Left is pony, right is original flux
struggling to get a wet lookcould be much wetter
>>101711777>>101711786Some guy already came up with an experimental training script to finetune Flux Dev. I think it's only a matter of time.Besides compute power, another potential factor that might hinder finetuning is the way Flux dev and Flux schnell were distilled from Flux pro. I don't have the full picture but apparently it's got something to do with the negative prompt and the guidance scale being hardcoded during the distillation process. However folks are already throwing out ideas to get around this limitation.
>>101711813>some random faggotThe CEO of black forest?
Yep. This is going to my insta account.
>>101711846Yeah, but it takes 75gb of VRAM to train a lora at the moment. Let's see if it can go down to 24.
>>101711861catbox?
Depict a dynamic and charismatic Hatsune Miku standing in a Pokemon stadium. She is shaking the hand of Ash Ketchum from the anime Pokemon. The image is in the style of the anime Pokemon. The scene is dynamic and features a Pokemon battle arena.
>>101711878I love how powerful the prompt "in the style of ____" works in flux
>>101711869I don't think that's gonna be possible without significant concessions (which would be detrimental to the final LoRA quality), or a new and significantly optimized LoRA training method.
>>101711875https://files.catbox.moe/xivwep.pnggotta use controlnet on this bad boy
>>101711857You are dumb. That was Invoke CEO
>>101711930baka I thought you were using flux
>>101711932That nigger also said that Inpainting was going to be impossible in Flux and a few hours later it was already working KEK!
I'm using the dev version of FLUX, but every seveal gens, it gets REALLY slow. I have to restart ComfyUI to get it working. What could be the issue?
>>101711936I'm really impressed at how good Flux is at knowing when something should or should not be symmetrical.
Does anyone know if Differential Diffusion amplifies Flux's inpainting ability? Or is it an SD-only method?
>>101711951of course not, fuck new things.
>>101712006based
>>101711963probably a leak in one of those shitty python packages
Anyone else recall when SDXL was released and people were saying it would be impossible to train a LORA on a 24gb gpu?
I am 100% convinced it's over for womancels. 5 more years of llm development + embodiment and it's over. If you think the current birth rates are low, it'll be fucking all over in 5 years.
>>101712095Yes, and I ended I trained a bunch for XL on my 2060S
>>101711891but this looks like the generic flux anime style and not pokemon.
Black Forest is promising text to video now. Get hyped.
>>101712161if you specify pixar its an entirely different result, it knows styles
>>101711535Thanks for the prompt anon
>>101712172why, I doubt I have the hardware to use it
>>101712174eeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee
>>101711718Very disappointed, yes, but on the bright side I can create training data for SD using Flux gens.
has anyone tried img2img with flux?
>>101712239It can mix styles? That's impressive. Prompt?
>>101712255credit goes to the guy in the /lmg/ thread, but you can specify different styles by saying the thing is "edited in">circa 2015 poor quality mirror selfie taken with iPhone in bathroom of overweight man with a greasy shirt and thick glasses with his arm over the shoulder of a cutout of Sailor Moon which has been edited into the photograph, Sailor Moon is blushing and the man is looking into the camera with a smug grin
>>101712253whats SAI doing right now anyway? havent followed anything, are they still begging for money?
>>101712312They said they won't be releasing any other SD 3 model and they are still confident SD3 medium will become the standard and Flux will flop because it cannot be fine tuned.
>>101712312they are just searching for a small loan of 30m dollars as this is the only thing preventing them from the release of the totally existing Stable Diffusion (NON BETA) Version 3.1: Awesome Edition which will bring world peace and answer the question if there are any odd perfect numbers at all existing (an unsolved mathematical question)
verify i am human.. but how....
Are those pants even legal?
>>101712431How do you folks get such cool poses?
>>101712431How did you get those pants, anon? I want to buy some for my gf.
>>101712103womancels?
>>101712469I was trying to be funny as if women were getting replaced soon, but came across as incoherent sincere schizo
>>101712448controlnet depth midas with sdxl. Lower the control weight to about 7 (gotta experiment with those).>>101712452after doing what I said above, you send it to img2img and just roll with cfg at around 0.5
Between Pony models and SDXL anime models, which ones are better for prompt adherence? Lets say I want to prompt a specific pose without using controlnets. Which models would give me a better chance of nailing it in my gens?
>>101712387i dont care as long as trani sufferslets hope the company dies really slow and that at any point there is a slight hope for it to be redeemed (which it will never be)that is really good for your mental health
>>101712494>controlnet depth midas with sdxl. Lower the control weight to about 7 (gotta experiment with those).Ah, so I imagine you would be using a reference image for the pose right?
>>101712499pony anime checkpoints usually dont need controlnets, but you can get any pose you want if you use a reference photo and use controlnet openpose on it.controlnet union is an all in one SDXL controlnet model btw, it works for canny/depth/openpose, works really wellhttps://huggingface.co/xinsir/controlnet-union-sdxl-1.0/tree/mainjust rename the promax model to controlnetUnion.safetensors and put it in the controlnet model folder.
>>101712514Yes. Jojo's bizarre adventure is your source for cringe poses.
>>101712539Thanks. Lets say I want to save a bunch of pose depth maps and openpose keypoints as images for later use. When I use these reference images, do I need to resize them to be the same size as my latent?
>>101712572if you want to save the images of the openpose stuff you can use something like this:https://github.com/fkunn1326/openpose-editoror, just save the image that has the pose you want and drop it into controlnet when you want that pose.
>>101712600Yeah but lets say I downloaded a depth map from somewhere and it's some random resolution like 1118x956 and my latent resolution is 896x1152. Do I need to resize the depth map manually before passing it to the Apply Controlnet node? (I use ComfyUI)
>>101712629i'd use a new depth map made in controlnet every time, otherwise the info might not work well for the new image: controlnet can make it fast anyway
>>101712312they basically got omega cucked by BFL releasing flux like they did. BFL released an 'open' model to drum up hype and attention for their 'pro' saas offering. that was basically the exact play that SAI was angling for. they were aiming to release SD3.1M with a goal of driving people towards their saas ecosystem, which would have their SD3L plus all their other random stuffso what happens when SAI drops SD3.1M and it looks retarded compared to flux? they're basically competing for the same audience (if there is one)
>>101712638That's not my question tho. I'm asking, if I don't have the original image and just the depth map, would I need to resize it manually to be the same size as my latent?The reason I'm asking this is cuz certain reference images may have other artefacts (like a table in the background, or another subject) which I may not want. So I might choose to generate the depth map, manually remove the unneeded artefacts from the depth map in an image editing software, and save the depth map with only the subject to be used later.Now, when I import this depth map later, do I need to resize it manually to be the same size as my latent?
>>101712660nta but you dont have to resize if it's not much bigger than 1024x1024 (I wouldn't resize from the example you gave). I have 16gb vram and haven't had any problems with those dimensions.
>>101712660I dont think you need to resize it, whatever you generate will be based on that map, the output is whatever size you like. So its up to you if you want it a rectangle shape, square, or whatever: but the initial output will be 1:1 with the depth map you pick, how it appears (skewed or not) depends on your output size.
https://suno.com/song/bbc7032a-93a5-445b-a0d5-c91563c5214d
>>101712685I see, thanks
>>101712710like if I did a canny controlnet on a 500x900 image, and the output was 250x250, the output will be fine but the output size will squish it. or you could also try the "resize and fit" option.
"cutout of _____" works great to mix real and anime, good job prompt anon
>>101712739Amazing.
>>101712755it's neat how dalle-3 is closed source and now we got an open source dalle, essentiallyor 1.5/sdxl/ponyxl checkpoints and loras for anime/realism gens as well
>>101712796>>101712739NINTENDO!!!!
>>101712801how did you get a realistic body but still get anime miku?
>>101712790prompt?
>>101712796are there /v/ threads for flux?
>>101712844not sure, but i'd prefer to see new flux/SD gens only in one thread so this is still the best spot.
>>101712844The dall-e thread is the designated Ai thread. Most anons won't care what model you use.
>>101712864Lol Are you doing something like cardboard cutout/Promotional cardboard cutout?
Does anyone have an install guide?
>>101712868I'm kind of surprised. 4chan is rife with in-fighting regardless of which board you go to. I figured /v/ would be pretty quick to jump on the "anyone still using dalle is a m$ dick sucker">>101712908https://comfyanonymous.github.io/ComfyUI_examples/flux/
>>101712908Be more vague please.
>>101712903another anon noted a cutout worked for putting anime into real life photos, so this includes "a cutout of Hatsune Miku wearing her original outfit which has been edited into the photograph"
>update comfy>fails>fresh install>install missing nodes>works>restart comfy>this error every timeWow really fucking cool
>>101712925Update Impact Pack
Filled!
>>101712925>comfy>it's not comfy
>>101712935Altho, I was going through the Impact-Pack main branch earlier today and I literally could not find a definition for the UltralyticsDetectorProvider class. Not 100% sure about this but it's probably been renamed to mmdet or something.
>>101712935Same thing, after update works but same error once I restart
>>101712913/v/ its full of people that aren't tech savy so things like online image gen gets far more use than local from what I've seen. Also plenty of people that have consoles and maybe a laptop but no dedicated pc.You can always make your own thread and see how it goes. If there was a thread with instructions on how to install local I bet it would catch on more.
Please someone bake
>>101712995>You can always make your own thread and see how it goesoh I don't actually care at all. I was just curious
>>101713010no, u
Next thread>>101713099>>101713099>>101713099Have fun ;)
>>101710957Has anyone tested this yet? Not at computer but the GitHub sample images don't explain much to me, would be nice to see a normal user xy
>>101711806if your buckets are different resolutions with different amounts of pixels (width*height) then buckets with more pixels will take longer
