Previous /sdg/ thread : >>101664139>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
>mfw Resource news08/01/2024>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>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>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>Accelerating Image Super-Resolution Networks with Pixel-Level Classification https://github.com/3587jjh/PCSR>ComfyStereo: port of the stereoscopic script used in stable-diffusion-webui-depthmap-scripthttps://github.com/Dobidop/ComfyStereo>MeGA: Hybrid Mesh-Gaussian Head Avatar for High-Fidelity Rendering and Head Editinghttps://github.com/conallwang/MeGA07/31/2024>Bubble Prompter for Stable Diffusion WebUIhttps://github.com/captainzero93/sd-webui-bubble-prompter>Waifu Diffusion V public testshttps://huggingface.co/waifu-diffusion/wdv-tests>ComfyUI_frontend v1.2.7https://github.com/Comfy-Org/ComfyUI_frontend/releases/tag/v1.2.7
>mfw Research news08/01/2024>Tora: Trajectory-oriented Diffusion Transformer for Video Generationhttps://ali-videoai.github.io/tora_video/>Dynamic Object Queries for Transformer-based Incremental Object Detectionhttps://arxiv.org/abs/2407.21687>Expressive Whole-Body 3D Gaussian Avatar https://mks0601.github.io/ExAvatar/>MTA-CLIP: Language-Guided Semantic Segmentation with Mask-Text Alignmenthttps://arxiv.org/abs/2407.21654>Evaluating SAM2's Role in Camouflaged Object Detection: From SAM to SAM2https://arxiv.org/abs/2407.21596>A Simple Low-bit Quantization Framework for Video Snapshot Compressive Imaginghttps://arxiv.org/abs/2407.21517>Fine-gained Zero-shot Video Samplinghttps://densechen.github.io/zss>Generalized Tampered Scene Text Detection in the era of Generative AIhttps://arxiv.org/abs/2407.21422>Benchmarking AIGC Video Quality Assessment: A Dataset and Unified Modelhttps://arxiv.org/abs/2407.21408>SmileyNet -- Towards the Prediction of the Lottery by Reading Tea Leaves with AIhttps://arxiv.org/abs/2407.21385>AI Safety in Practice: Enhancing Adversarial Robustness in Multimodal Image Captioninghttps://arxiv.org/abs/2407.21174>Optical Computing for Deep Neural Network Acceleration: Foundations, Recent Developments, and Emerging Directionshttps://arxiv.org/abs/2407.21184>Embedding Space Selection for Detecting Memorization and Fingerprinting in Generative Modelshttps://arxiv.org/abs/2407.21159>Adding Multi-modal Controls to Whole-body Human Motion Generationhttps://yxbian23.github.io/ControlMM/>Direct Unlearning Optimization for Robust and Safe Text-to-Image Modelshttps://arxiv.org/abs/2407.21035>Safeguard Text-to-Image Diffusion Models with Human Feedback Inversionhttps://arxiv.org/abs/2407.21032>SuperVINS: A visual-inertial SLAM framework integrated deep learning featureshttps://arxiv.org/abs/2407.21348
First in thread. What's up chaps.
>>101670503>Would berry my dick so far in her ass
>>101670535>Error: Our system thinks your post is spam. Please reformat and try again.for some reason, I can't post this in the news because 4chan is flagging it as spam. lets see if it can be a standalone post>ComfyUI: Basic Flux Schnell and Dev model implementationgithub.com/comfyanonymous/ComfyUI/ [COMMIT ID GOES HERE BUT IS SPAM APPARENTLY]
anyone tried flux yet? Should I care?I have 24gb vram
>screenshot from a brazzers.com production
>>101670570if you have a 40GByte VRAM gpu you can go download a new 12b parameter model>https://huggingface.co/black-forest-labs/FLUX.1-dev>https://blackforestlabs.ai/announcing-black-forest-labs/and run it on comfyif not.. same old sam old.. SAI still has their head in their ass
FLUX-devhttps://replicate.com/black-forest-labs/flux-devFLUX-schnell (distilled)https://huggingface.co/spaces/black-forest-labs/FLUX.1-schnell
>>101670606looks like the examples are from saas and local is resource heavy and not the same >>101670081 >>101670180
>>101670606no.. the model alone is 24GB
What's the deal with SD3? Last I checked in here, people were saying>Garbage output>Trash liscenceStill the case? Anyone using it?
>>101670582Links which have too many slashes elicit automatic spam protection.
>>1016706692 more weeks for the fixed version
>>101670668What even is this
>>101670669the beta model is out but the 3.1 is still slated for release
>>101670606it's quite good >>101670319
>>101670582You can use a 6 characters commit id in the link instead.https://github.com/comfyanonymous/ComfyUI/commit/1589b5
>>101670683Why even bother posting
>>101670616Can it do celebrities?
I miss PW, he is a tarento.
>>101670701great, thanks! I'll use that for the next news post
>samurai bunkek
>>101670775In Japan, this would mean the Samurai was in debt for someone. Bartender would cut off his bun and sell it back to him.
8 steps flux dev206 seconds
>>101670807ouch.. how much vram and ram did it gobble?
Can anyone please put the flux model in a torrent?
>>101670723Audrey Hepburn8 steps
This model is pretty goodhttps://files.catbox.moe/79tj08.png
>>101670830just make a huggingface account, the dev version is free
>>101670807stop the shilling, so far no "flux" gen has been better than any sd 1.5 or xl gen
https://files.catbox.moe/gej765.jpg
>>101670941best ass ever. moreI want a gf who looks like that so bad.
>>101670918if you're not a sai employee, idk why you'd care what models people use
>>101670864wtf those nipples are true mosquito bites lmao
>>101670736>he
>>101670918weirdo
https://replicate.com/black-forest-labs/flux-dev>Great image quality>Ok prompt understanding>Can do NFSW>>101670941>>101670424 >>101670391 >Nice anatomy>Apache 2.0 LicenceThis is a great day, fuck you SAI
>>101670642>no.. the model alone is 24GBlooks like it can be run on a 24gb vram card >>101670797
>>101671037>too big to run locally unless you're GPU richforgot that bit
>>101671114you can't make a great model if it's not big enough, you also forgot that bit
>>101671143now local can make meme images and clipart like dalle!!111!!
>>101671037>Can't generated a guy doing a handstandThere is still room for improvement. It also can't generate broken objects, like a broken sink or window. I also tried the "woman stepping on a cake" test; the cake looks as hard as rock. When are we going to have a physics-aware diffusion model?
The Schnell demo has Fallen!
>>101671143I consider "can I run it" a component of whether a model is 'great' or not
>>101671180retarded faggot
>>101671158it has great anatomy, great image quality, insane prompt understanding, the fuck do you want more?
>>101671180but it's possible, if you quantize to 8bit, and DiT models are known to be really resiliant to quantizations because it's a transformers model
>>101671205goes without saying
>>101671205you new?
>>101671214for it not to take forever? For it not to use the wasteful t5xxl instead of t5xl? it can only make images and there is no room for vid tech that isn't txt2vid boring garbage?
The Schnell hugginface demo Is back!
>>101671258>it can only make images and there is no room for vid tech that isn't txt2vid boring garbage?Moving the goalpost I see.>Yeah messi is great at football... BUT NOT SO GREAT AT BASKETBALLwe never asked this model to do that faggot
>>101671231sure, I'll look forward to that but for now its too big for me
>>101671293use that instead anonhttps://replicate.com/black-forest-labs/flux-dev
>>101671301your uncle was too big for you too
>>101671300>Moving the goalpost I see.these are problems with the model and you don't have a comeback for it>clueless comparisongo sit in the corner retard
>>101671329>these are problems with the model and you don't have a comeback for itI never said this model is perfect, why are you lying like that? I said that it's by far the best base local model we ever had, it's not even close
>>101671312Kek, Lego Quokka turned into duplo fox
i miss schizo anon
Flux is incredibly good at text, and real text, not the photoshopped comic--sans text like SD3 kek
>>101671415I'm right here.
>>101671341>lyingpoint out the lie
>>101671416please do a proper upscale of the bottom left one, very nice, or hand it over so I can do the dirty deed
>>101671499the image there is at its right resolution, you just have to crop it and upscle
>>101671499Is it even possible to upscale in comfy if the normal gen eats exactly 24gb of vram?
>>101671558it won't be 24gb of vram if you load everything on 8bit anon>--fp8_e5m2-text-enc --fp8_e5m2-unet
>>101670535Listen carefully, I'm not going to repeat:DOWNLOAD FLUX SCHNELL FAST!!!!!!!!!! It's probably going to get taken down or some shit.The niggas didn't filter the dataset properly, it can generate realistic looking ch*ld nudity. Right now Replicate/Fal don't filter the model outputs. They'll notice sooner or later. IDK how they did such a bad job - the anime seems to be filtered from NSFW, but not real world fucking photos.
>>101671114>hardware for a 12B model>gpu rich/lmg/ with their 4x4090 rigs would laugh at you
>>101671579>DOWNLOAD FLUX SCHNELL FAST!!!!!!!!!!more like flux dev, that's the better onehttps://huggingface.co/camenduru/FLUX.1-dev/tree/main
>>101671579>The niggas didn't filter the dataset properly, it can generate realistic looking ch*ld nudityretard fearmongering for no absolutely no reason
>>101671626it's d*bo
>>101671616yeah,, I checked, dev can generate it too, but schenll does it in literally 1 second on replicate
>>101671579Since when it's a problem, retard? All SD models could generate that. And SD 1.5 is especially great at it.
download dev + schnell fast!!!!!!
>>101671645You haven't seen schnell outputs, they're complete level above SD 1.5, idk what dev could do given the right prompts.
>>101671656Niggy, that was the very first thing I genned and it wasn't as good and hot as what SD1.5 or SDXL produce.
>>101671579omg, its unsafe? I'm writing my congress persons as we speak>>101671597all that GPU just to make a miku chatbot>>101671645>Since when it's a problemcuz they'll get sued and have to spend all their money on lawyers instead of training models
is animatediff dead? i see the motion model for xl is 9 months old and is a beta. how are people making animations with pony or xl?
>>101671656>>101671647>>101671579that's cool, the fearmongering level is a good metric to see if a model is usually good or not, unironically
>>101671037not bad
It passes the grass test alright
So what's the soft ceiling for good quality with SDXL? It seems that when I start genning 1440pixel images, the quality starts dropping compared to like 1000pixel. Is there a sweet spot?
>>101671726It perfectly passes the grass test, insane qualityhttps://replicate.com/black-forest-labs/flux-dev
fp8 loading just dropped in comfy
>>101671774>unet_nameso flux is a unet architecture?
>>101671416
>>101671788it's insane how those humans look at far away distance, more parameters is really all you need after all
hmm... it's a good effort I guess
>>101671037>Apache 2.0 Licenceproprietary license. only the vae is apache 2.0
the anime is pretty good. pro tip: object placement is always from the viewer's POV. I tried doing it from the character's but it always favors the viewer
For those who have at least 16gb of vram, there's a script that can run flux at 8bithttps://gist.github.com/AmericanPresidentJimmyCarter/873985638e1f3541ba8b00137e7dacd9
>>101671823Damn, even an American President is using Flux lol
>>101671799ehm?
>julien
>>101671815>proprietary license. only the vae is apache 2.0is it worse than SDXL licence? I hope not, pony needs to work on this shit!!
>>101671037I'm sorry but in >>101670941 the girl's left hand is hairy.
>>101671856>implying a hobbyist can afford to train it
>>101671853it's way better than any other model at this far distance anon, not even close
>lubimiv
I showed flux to my gf and she said it's deeply concerning because people would use it to make nudes without consent
>>101671876hobbyists train 70B local LLMs, surely they can train this
>>101671823i thought you built that 4x3090 rig
What is the point of SDG and LDG separation if everyone uses flux?Asking for future.
>>101671882so she's a glass half empty type of person
>>101671897>everyone
>>101671882It really is disturbing that people make nudes of real people when you can make nudes of far hotter fictional people
>>101671887they are using 130b param models and have room to spare yet we are maxxing out at 12b in imagen... they shouldn't be treated the same
>>101671906Like the celeb nude creators that just ruin everything for everyone.
>>101671897I'm not getting a 3/4090 until they're £250~maybe the bubble with have burst by then and this is the peak
>>101671897/ldg/ is full of losers who bet on lamer models than what sai shit out. most of the people from BF are ex-sai shitters too
>>101671906this, it's fucking retarded
>>101671906it's also a crime to make nudes of people without their explicit consent and downright repulsive
>>101671799dude thats an upscale with an nmkd model, an sdxl model resample and a facedetailer pass. adjusted the detection threshold so it picked them all up, even the 2 in the middle. not too much denoise so they dont stick out too much.
>>101671906>>101671882nice, the fearmongering level is quite high, that means it's a great model >>101671626
>>101671579It's time someone simply stated: "we don't ban photoshop or cameras because they can be used to make illegal pictures, AI models can generate illegal pictures especially by a determined user, they are accepting that if they're caught with those illegal pictures they will go to prison"
>>101671978say that to the jews
>>101671980People only censor models because of the perceived social backlash.
>>101671897/ldg/ split off because they wanted a thread that was safe for trannies, not because they had any better models to use. /sdg/ primarily uses the best model, which until today was SD versions and their finetunes, but other models have always been discussed and occasionally posted with alongside them without issue.
>>101671996The split is because SDG attracts autistic weirdos that treat 4chan like a Discord server.
cant quite do catgirls of my preference
dall-e 4 when
>>101671995and abetting illegal counterfeiting.
>>101672033flux v2 will be dalle4
>>101672014/sdg/ is where the cool kids hang out 8)
>>101672019I can't get anthropomorphic dog people competing at the olympics either>>101672074I wonder what hes thinking rn
>>101672084yeah /sdg/ loves kids
>>101672019>ComfyUIare you using it on fp8 anon?
>Debo still posting blurry gens even with a new model Pottery
>>10167201910/10 would pat
>>101672102there's nothing wrong with liking kids, having a sexual attraction for them on the other hand....
>>101672115i just have the "weight dtype" set to default, didnt know what it did so didnt touch it
howifeel
>>101672019>cant quite do catgirls of my preferenceskill issue, unironicallyhttps://replicate.com/black-forest-labs/flux-dev
>>101672188How much VRAM do you have?How much VRAM does it ask for the image model?How much VRAM does it ask for the text encoder model? Do you offload?
>>101672217i have a 4090, i dont do anything that i know ofim running on wsl if that mattersloading in lowvram mode 21557.27999973297i get this log when running which i've not seen before
>>101671726FP8 might be degrading quality in the fine details?dunno
Flux is greatYou need a 3090/4090 for it thoughWith the basic workflow from comfy it fills up to 23.5G in lowvram mode30 seconds/20 steps on my 4090but quality is topwe gonna eat goodthis feels like back when 1.4 was released
>>101672094
>>101672303>With the basic workflow from comfy it fills up to 23.5G in lowvram modewhat does lowvram do? it offload the text encoder in the cpu?
>>101672323might be easier with cartoon styles. I'm trying to make it a real tv broadcast (but failing on the aesthetic too)
>>101672303>You need a 3090/4090 for itit's over.
>>101672303yea just got it working to .. but it lets test its limits, btw 1072x1072 just barely works on 24GB to
>>101672367it's still better than what the /lmg/ fags have to deal with, they need multiple 3090 to be able to run decent models kek
>>101672375>>101672303>>101672279>>101672263can someone do a comparaison between fp8 and fp16 to see if there's a lot of difference? if not you can save a lot of vram
>>101672388aren't you an /lmg/ fag?
>>101672388now you need it too, got back in the closet.
>>101672397Looks pretty nice actually
>>101672367hang in there, quantized model in 2 weeks
>>101672409he is, /lmg/fags are known for their shit tastes & miku avatarfagging.
>>101672409>>101672411I am and I have multiple gpus yeah, unfortunately having multiple gpus don't mean shit in image gen because you can only use one
oh yeahit's good
https://huggingface.co/camenduru/FLUX.1-dev/tree/mainDo I have to rename .sft to .safetensor to make it work on comfy?
>>101672404can't for fp16 you need something with more than 24GB of VRAM
>>101672463surprised what characters it actually knows. Seems like several were aligned out (asuka for one).
>>101672475no need to rename itjust put it in models/unetand put the ae.stf in models/vae
>>101672479>can't for fp16 you need something with more than 24GB of VRAMsome anon managed to do it on fp16 though? >>101672263
>>101672509what about the text encoder?
>>101672540goes in clip
>>101672560I put the three of them?clip_l.safetensorst5xxl_fp16.safetensorst5xxl_fp8_e4m3fn.safetensors
>>101672019adorable regardless
Aw hell yeah!
wowworks well on resolutions lower than 1024x1024I think they are using Unet for this type of versatilityDiT would usually shit the bed when trying to generate on resolutions different than what it was trained on
>>101672581>works well on resolutions lower than 1024x1024indeed, the blog says that it should work fine on low resolution, they are so based they thought of everythinghttps://blackforestlabs.ai/announcing-black-forest-labs/
>>101672574they all go intol >/models/clip
>>101672279>>101672404okay FP8 is a tradeoffthe degradation is subtle thoughalso, someone needs to figure out how to hi rez with this model because the standard Img2Img is causing the colours to go whack>>101672579nice>>101672581Flux is a Mixed DiT model
>>101672610thanks a lot anon!
can do goo, cant feet>>101672610
>>101672479>>101672522I have now swapped to fp8 as it was occasionally getting this, and it hasn't happend sinceLoading 1 new modelPrompt executed in 35.02 secondsgot promptmodel_type FLUXKilled>>101672404"a robot chicken"
>>101672428python is really shit trying to make that work while all the llm repos are in cpp
>>101672623okay fp8 TE was the culpritfp16 TE with fp8 DiT is close enough
fp8 on a 4080, 1.21s/it
>>101672623for the time being you can hires with sdxl if needed.. first I have to find benefits this models offers other than extreme text accuracy and way more knowledge since 12b ofc .. gotta poke it with the go'ey stick>>101672660I tried fp16 and got a kernel panic and blue screen, glorious
>>101672690how do you set fp8 on comfy?
make a huge titty bimbo in flux, right now.
>>101672692that's not bad
>>101672743woa have your read the license? violence against children in gens is forbidden
>>101672708git pull and get the latest Comfyuse these example workflowshttps://comfyanonymous.github.io/ComfyUI_examples/flux/set weight_dtype to one of the fp8 in Load Diffusion Model node>>101672702cute slimes
>>101672279>>101672690Thanks for the feedback anon
>>101672751she's not a real person
>>101672752>cute slimesthank you.. I prefer my sd slimes sofar.. also it doesnt understand "transparent slime"
the quality is extremely impressiveif you told me this was generated I would not believe you
>>101672752>set weight_dtype to one of the fp8 in Load Diffusion Model nodewhich one? fp8_e4m3fn or fp8_e5m2
upscales well
I just realized that it has no negative prompt
>>101672755uhh you got them wrong anonplease delete your image because it is misleading, I'll post a comparison pic later>>101672802either should be fine, I used fp8_e4m3fn
>>101672794>the quality is extremely impressive>if you told me this was generated I would not believe youthe very first flux gen I saw was this one and I thought it was just one of those random memes you found on the internet, that was a great first impression yeah
>>101672829what do you mean? I put his fp8 as a comparaison? it's not the good one? >>101672690
>high quality movie poster of Elsa as John Wickfuuuuckit's so goood
Can any one replicate this style
>>101672853all those pics are minethey are not a proper comparison because it is a mix of fp16 and fp8 TE as well as DiTI'll create a proper comparison of bf16 vs fp8 DiT with TE fixed to fp16
>>101672855>>101672883which model version? can you pls share, since flux has 3 different ones
>>101672870gave it a try>high quality sketch illustration of a badland female warrior with tattoos. her hair is a mess and she is eager to fightnot quite there, but with a few tweaks it can get there
>>101672894>it is a mix of fp16 and fp8 TE as well as DiTflux is a DiT model?
>come out of fucking nowhere and btfo everyone in an instanthow did they do it
>>101672904I'm just on flux-dev for nowhaven't tried the other one yet
ya 12b is indeed capable of nice stuff>a complex science fiction town drawn by Akira Toriyama, sunset .. now I wish you could get an anime nipple, wasnt able to yet
uh, lewd huge tits?
>>101672941is that flux?
sdxl, 1min47sex, gtx1060 6GB, 16GB ram
>>101672953paying her respects for his parents :(
>>101672941ah, wrong one>>101672956sdxl
>>101672918
FUCKING KEK, STABILITY IS D E A DEAD
>D'Oh!
how do I put the text encoder in fp8?
>>101672961same min same sex
guess we will have to wait a few months till a capable anime fine tune of flux appears .. its okaish in anime, nothing more .. or I havent found the right way to prompt it into good anime yet
>trani
wtf, flux asks for a shit ton of ram during loading
>>101673001top right corner, change "weight type" from default to fp8_e4... or fp8_e5...
>>101673001set weight_dtype to one of the fp8 in Load Diffusion Model node>>101673024what are you looking for anonSFW anime should be okay in Flux
>>101673044>>101673045but that will put fp8 on the image model and not the text encoder? or it does both? how do you get to choose only one of them?
working on flux ait desui really need to fix up more of ait so its easier to write the IR. the rope stuff is just awkward atm
>>101673045>SFW anime should be okay in Fluxhhhmmm thats not what I mean .. it doesnt understand some basic stuff I am used to .. like transparency on 2D .. or arists .. maybe I just need to learn to use txxl5 betterpic related for example is _not_ an>anime girl by Studio Ghiblieven NAI got Studio Ghibli right .. probably we just need a finetune
>>101671926NINTENDOHIRE THIS COMPUTER
>>101672982So Flux is actually the real SD3. We're so fucking back
>>101672428>multiple gpus don't mean shit in image gen because you can only use oneWhat, really?There goes my 2x3090 idea on this.
Cool
>I'm sorry SAI. Nothing personal. It's just ... business.
how much VRAM do I need for flux?
>>101673070you just download the fp8 t5 model and use that: https://huggingface.co/camenduru/FLUX.1-dev/tree/main
>>10167312232Gb
>>101673129and I sleep again
>>101673135works with 24GB in float8
>>101673070that should change only the diffusion modelfor TE fp8, just load the fp8 t5xxl file or add the corresponding cmd line option when running comfy>>101673096artists are probably nuked, and 2D data seems to be cleaned, so transparency might be an issuelet me see what I can gen
>>101673122you'll never be able to run that model, anon. start working some overtime shifts and you might get a chance next year
>>101673117
>>101673117>>101673162That's fucking crazy how accurate it gets the default XP wallpaper
>>101673147So basically I put fp8 for the clip 1 and I let clip_l for the clip 2? Is there a big difference between a fp16 text encoder and a fp8 text encoder?
>>10167312224 GB>>101672774try "transparent see-through"picrel is "a transparent see-through cute slime in a foggy moody forest"
>>101673093>aitthe more things change, the more they stay the same
>>101673070you are right just look at the left most nodes. .the top is the model, you can choose to go fp8 on that. the lower one is the text enconder, you can go fp8 there .. either will do, or both (maybe you can get higher res then)also pic related.. somewhat studio ghiblish.. hmm .. I just dont like boomerprompting
>>101673176it's really good at making desktops, give it a try>>101671850>>101671903>>101671959
>>101672909Its also the lighting and softness
lots of dreaming last night. only bit i remember is at the end a sort of music video for 'school's out' starts playing but it devolves into a highschool of zombie people, and a young boy gets led around the corner and torn up. then i was awake.
How you you go on preview mode on Comfy?
>>101673180>>101673178yesfp8 TE is slightly jank, better to use fp8 for the DiT since it degrades much less compared to the TE
>>101673180ya thats great, but thats not anime style, its more like a 3D render, you can achieve such things in realistic SDXL models to, but I am looking for something thats transparent in a 2D environment
>>101673210I meant within the context of it being in what is essentially the background.It's an unnecessary but impressive detail.
closer to Dall-e quality than anything we've had yet, crazy this dropped outta the blue
I am probably just gobly that I have to learn new prompting rules .. itll work out, 12b definitly is superior in many many ways
>>101673218hop on the flux train and you'll forget all about it
>>101673219i assume you mean like a "simple ui" view?drag the clip text encode above the picture and zoom in so it fills your screen.boom ui engineer
>>101673221>fp8 TE is slightly jank, better to use fp8 for the DiT since it degrades much less compared to the TEand the text encoder is on the VRAM right? is there a way to put it only on RAM so I can save memory for the DiT?Btw fp8 DiT managed to do the "halo" meme fine
>>101673196at this point there's so much more in my fork of ait and meta's i might as well defork it
>>101673258no, like you can see the steps as it starts to generate, like on A1111
>>101673219double click an empty areatype preview and click the Preview Image nodeconnect the output of the VAE Decode node to thattoggle the save image node with Ctrl + M>>101673228I get transparency in 2D just fine thoughpicrel is "anime illustration of a woman wearing a transparent see-through dress">>101673265Comfy will automatically shuffle things around between RAM and VRAM, don't worry about it
>>101673279open comfyUI manager menu and select a preview type in the dropdown box
>>101673270than meta's*
>>101671258They stated the next model is a T2V based on Flux.
>>101673239holy i am coom, give prompt pls pls
ram usage converting dev to diffusers lol
>>101673286>I get transparency in 2D just fine though>picrel is "anime illustration of a woman wearing a transparent see-through dress"ya I guess transparent anime dresses are abounded in the data set, but it seems neither are oppai girls nor transparent slimes>an oppai girl is hugging a transparent see-through slime in a foggy moody forest at sunset, kawaii, anime style, behind them are sticks and stonesnothing to worry about, finetunes will happen if the license allowes it I guess
>>101673279https://github.com/comfyanonymous/ComfyUI?tab=readme-ov-file#how-to-show-high-quality-previewsbut I think someone will need to write a latent2rgb/train a taesd equivalent for this model
>>101673286it understands 2D transparency just fine>>101673340It can do isometric GBA style, nice>>101673343lemme see if I can get something similar
>>101673296>open comfyUI manager menuwhere is it?
AAAAAAAAAAAAA FUCK ME replicate flux pro is broken
stability wasnt even able to release the updated SD3 weights lmao
>>101673384Whatever they release is gonna look like a pebble compared to the mountain Flux is
For each gen, it loads and unload on my ram, that makes shit slow, how to prevent that on comfyUI?
>>101673375oh, you need to download ComfyUI Manager. You can enable previews with startup commands too, but idk what they are, you'd have to look it up. ComfyUI Manager is incredibly useful tho, so if you want to actually start using comfy I'd recommend getting that: https://github.com/ltdrdata/ComfyUI-Manager
>>101673384instead they left stability and then released flux :^)
>>101673367>lemme see if I can get something similarthank you .. ill try to learn .. txxl5 gives me headaches "retro anime style" kinda works tho
does that count as transparent
>>101673384It's because the people who were originally making SD3 left to make this kek
>>101673399>>101673399>>101673399
>>101673435oh ok, thanks anon
>>101673445yea kinda.. but I guess I am just perfectionist with transparent slime.. thanks for trying tho
>>101673445If the cactus is behind her, yes.
>>101672982Yeah so the issue really was Emad. They are doing perfectly fine now that this grifter is out of the place.