Discussion of free and open source text-to-image modelsPrevious /ldg/ bread: >>101840282>Beginner UIEasyDiffusion: https://easydiffusion.github.ioFooocus: https://github.com/lllyasviel/fooocusMetastable: https://metastable.studio>Advanced UIAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiComfyUI: https://github.com/comfyanonymous/ComfyUIInvokeAI: https://github.com/invoke-ai/InvokeAISD.Next: https://github.com/vladmandic/automaticSwarmUI: https://github.com/mcmonkeyprojects/SwarmUI>Use a VAE if your images look washed outhttps://rentry.org/sdvae>Model Rankinghttps://imgsys.org/rankings>Models, LoRAs & traininghttps://civitai.comhttps://huggingface.cohttps://aitracker.arthttps://github.com/Nerogar/OneTrainerhttps://github.com/derrian-distro/LoRA_Easy_Training_Scripts>Fluxhttps://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellhttps://comfyanonymous.github.io/ComfyUI_examples/flux>Pixart Sigma & Hunyuan DIThttps://huggingface.co/spaces/PixArt-alpha/PixArt-Sigmahttps://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiThttps://huggingface.co/comfyanonymous/hunyuan_dit_comfyuiNodes: https://github.com/city96/ComfyUI_ExtraModels>Kolorshttps://gokaygokay-kolors.hf.spaceNodes: https://github.com/kijai/ComfyUI-KwaiKolorsWrapper>AuraFlowhttps://fal.ai/models/fal-ai/aura-flowhttps://huggingface.co/fal/AuraFlows>Index of guides and other toolshttps://rentry.org/sdg-linkhttps://rentry.org/rentrysd>GPU performancehttps://vladmandic.github.io/sd-extension-system-info/pages/benchmark.htmlhttps://docs.getgrist.com/3mjouqRSdkBY/sdperformance>Try online without registrationtxt2img: https://www.mage.spaceimg2img: https://huggingface.co/spaces/huggingface/diffuse-the-restsd3: https://huggingface.co/spaces/stabilityai/stable-diffusion-3-medium
Blessed thread of frenship
>mfw
>be me >cant run flux with 3080/10 GB>figure i should get 32GB system RAM>get 2x16 of the exact same brand/speed/voltage/etc as my current ram>order on Amazon>"same day delivery for FREE" why not>it arrives this evening>in a flimsy paper envelope>with no dunnage whatsoever>install it>fire up computer>after a few minutes, BSOD>restart a few times>BSOD, again and again>take out the new RAM>no more BSODwhat now anons? do I buy different RAM and try again? do you think its amazons sloppy delivery? i really dont have the time to diagnose it but what software/driver issue would cause BSOD if it isnt defective RAM?
>>101842967meant to say, an additional 32GB to put me at 64GB
>>101842960Oh, that's funny. That image is at 21 steps. This is with 50 steps.The prompt contained a bit where it said she had a giant sword, which I specified since it felt like the swords I was getting were too small. Maybe I should do higher steps as a default.
>>101842967Update your BIOS if you haven't. Especially if you're on a platform like AM5 the RAM stability sucked early on. If not, you'll probably want to do some manual ram tweaking.
>>101842967The second computer I built, I had double RAM failure. PC could not post. Was a bitch to diagnose, since I could not know if it was mobo, cpu, ram or psu.Had to use third party repair shop where they could test every component and they identified that both memory sticks were dead. Retailer almost did not believe that I could have such a shitty luck that both sticks could be dead, but eventually sent them to the manufacturer and I got new ones.I don't know why I told you this, since it does not help you at all, but now you know.
>>101842979>same day deliveryuh oh, you bought make Chinese ram that temu resellers store at amazon warehouses.
>>101842967You need to buy RAM as a kit, mixing and matching is often an issue, even if it's the same model, serial etc, it's not the same RAM.
>>101842967>i really dont have the time to diagnose it(Assuming you're on windows) run mdsched.exe (memory diagnostic scheduler) from System32. This will schedule a memory diagnostic to run the next time you boot your computer. if the memory is the problem this will find it. If the problem is a bad sector this can "turn off" that sector with firmware and boot normally with the rest of the good memory HOWEVER in that case file an RMA with amazon.
>>101843084Wtf, why are you posting my picture here?
>>101843016It's true, but when Flux gets IPAdapter it should fix the lack of style/artist knowledge totallyYou show the model an image in the style you're looking for and it just werksOnly times IPAdapter doesn't work for imparting style is when the model just doesn't know how to do that style, which isn't a problem Flux has, Flux's problem is just not knowing the NAMES for things. So IPAdapter will solve it
we gonna make it
>>101842979>an additional 32GB to put me at 64GBDafug? I have exactly the same config (3080, 32GB RAM) and I can run flux no problem, in fp8 mode at least.
>>101843319>3080>32GB RAMHow many of us are there kek
>>101842967people have been running with less specs...what exactly is the issue?considering 1080ti runs it
>>101842967Try lower the RAM speed in the bios. When using 4 RAM slots you might not have as much overhead.
>>101843367:^)
>>101843373Lower to what? It was 3200 mhz, i changed it to 1600 mhz and still got BSOD. I think I'll try different RAM and if the same thing happens, then I'll diagnose
>>101843319why use fp8? when you can use nf4 (almost better in every single way)
is it justified to spend the cost of a 3090 on vast.ai if i'm only going to play around with flux for a month max?
>>101843404just send that fake ram back and get some other
>>101843411if ur not going to play games, sure. but 3090 is pretty cheap especially second hand. would honestly spend a bit more for a 40 series card instead.
Will Purple Smart Ai fix the score nonsense with his auraflow model?
>>101843404If your cpu has it you can try switch from gear 1 to gear 2 on the off chance it's the memory controller though what >>101843420 says is probably the best call.
>>101843429he talking about renting usage of one while he sits on his chromebook
>>101843319I was running Flux dev at 8, only worked if I allocate 32 GB of my SSD to paging file and it spikes to 99 percent usage when I'm generating Flux. I like to able to surf the web and multitask while doing gens
>>101843454If it's a flash-based SSD please be aware of burn-out
>>101843470Thats what I mean, I'd rather get more RAM then rely on paging. Not even gonna fuck with it
>>101843454have you tried nf4? i'd sure be using it if my gpu supported
>NF4 lets me fit Gemma2 9b and Flux in VRAM so I don't even have to prompt my own slopFeelsgoodman
Has anyone managed to make a realistic person with a very small head? I tried "size of walnut" and got picrel, but otherwise it's just dudes with normal sized heads
>>101843555nice nipples for a flux gen ngl
>>101843586Positive only anon here. I conceded btw.
>>101843629Welcome to prompting 101 dip shit
>>101843629I accept and respect your concession, not everyone is willing to admit they are wrong on the internet, so no hard feelings on my part and I'm glad you've learned something cool to improve your craft anon.
>trump smoking weed with snoop dogit's that easy huh
>>101843667I am still not going to use them because I can't stand how slow it makes my gens. I lose literally half to 3x the speed. Maybe if improvements can be made somewhere but who the fuck knows.
>>101843716Why is his hand white?
>>101843716>I lose literally half to 3x the speed. Maybe if improvements can be made somewhere but who the fuck knows.CFG > 1 is at worst 2x slower, not 3x. And you can really improve that speed if you use the AdaptiveGuidance threshold and put a low threshold, and it also removes the artifact the high CFG can make, it's really a win win situationhttps://imgsli.com/Mjg1OTU5https://imgsli.com/Mjg2MDc4https://reddit.com/r/StableDiffusion/comments/1enxcek/improve_the_inference_speed_by_25_at_cfg_1_for/And for realistic shit, AdaptiveGuidance starts to activate pretty quickly so you get some speed that is just slightly lower than a full CFG 1
Here's a schizo theory. Isn't it weird that flux can do virtually no celebrities except for trump, who it nails perfectly? What if this is intentional? They deliberately didn't clean him out with the hopes that some users would create incriminating material of cheetoh hitler. They're eurocucks after all, maybe they see it as some kind of noble humanitarian move.
>>101843818>They deliberately didn't clean him out with the hopes that some users would create incriminating material of cheetoh hitler.The opposite can be true, people can use flux to make cool trump montages
>>101843842There have been significantly more Tump doing cool shit images than incriminating ones.
>>101843902this
>emmawats0n eating ice cream with Donald Trumpkek, why do the trigger words on Loras can't be simply the actual name, going for "Emma Watson" instead of "emmawats0n"? It would even be better because flux knows Emma a bit, so the lora would improve on that instead of reinventing the wheelhttps://civitai.com/models/639737/emma-watson-lora?modelVersionId=715459
>>101843555>activated
>>101843802Do you mind posting your workflow? On my 3090 32gb ram system cfg > 1 results in generation times between 3-4 minutes. This is at 20 steps with no post processing.When cfg = 1, it is usually around 60 seconds.
>>101843920That's impressive how well loras are working on flux
>>101843842no gen can top the post assassination picture tho
>>101843974>Do you mind posting your workflow?Sure thing anon, here it is: https://files.catbox.moe/nz6p6t.png
>>101843978true, this is one of the most badass picture I've ever seen in my life, we're not gonna get something this good in our lifetime again
>>101844021>basedfacing so hard rn
>>101843998giwtwm
>>101843920You cant make her do unsafe things though. And that's not Trump.
>>101844115>You cant make her do unsafe things though.True, I'll be waiting for a porn finetune of flux, it's gonna be great
>>101844006Don't you need some kind of python script for this to work?
>>101844130Just follow that tutorial:https://reddit.com/r/StableDiffusion/comments/1enxcek/improve_the_inference_speed_by_25_at_cfg_1_for/Unless you're talking about loading the workflow? In that case you just go to ComfyUi, you click on "Load" and you load the picture
>>101843920maybe the maker was assuming theres some kind of textual censorship affecting the name in its correct spelling (which i doubt)
>>101843998That rabbiteer is so close to NSFW territory that his leg broke.
>>101844033>>101844026>>101844115>>101844121>>>/h/hdg>>>/e/edg>>>/b/degen>>>/d/ddg>>>/aco/sdg>>>/trash/sdg
Is "base_shift" from ModelSamplingFlux supposed to do anything? I changed that value a lot and nothing changes in my image
>>101844203that slides the woke affinity, go too far and you'll begin to transition
>>101844144have. All my outputs are white
Does (prompt:1.33) etc do anything for prompt strength on Flux
>>101844219most likely your dynamic thresholding node is out of date
>>101844219Update ComfyUi anon, it'll update the packages aswell
forge, nf4, gtx1060 6gb vram, 16gb ram
I like NF4 for the much more flexible workflow possibilities it opens up, but the idea that it's better than FP8 seems to be insane cope.Left to right is NF4, FP8 and FP16. Only NF4 has fucked up the child rabbit by giving it an extra hand on one side and getting confused about whether it's standing on the table or behind it.
>>101844269Forget to add in case it matters, T5 text encoder is in FP16 for all three.
>>101844269Yeah it is a cope, but that's impressive how well nf4 performs, not even a year ago you would consider the possibility of a 4bit image model you would be laughed at, and now we need to do some meticulous testings to notice big difference between nf4 and fp16, desu I'd love a nf8, that shit would be virtually the same as fp16 instead of staying on the outdated fp8 architecture
>>101843920>>101843977I have no desire to see a return of Watson posting but this is quite impressive for being only trained on 25 images
>>101844260that prompt format is...... suboptimal.>>101844302>not even a year ago you would consider the possibility of a 4bit image model you would be laughed atno, there has been studies on it earlier than that.
>>101844302Yeah as I said I do like it, because with NF4 I can now have workflows that load an SDXL model with controlnet etc. after the Flux gen to do stuff to it. There's lots of stuff where the loss of precision doesn't matter. I was just irked by the claim that it's more accurate than FP8.
>>101844269cute rabbits anon
>>101844342They're not on flux, so aside from the lack of score_x and not using 1girl instead of one girl it's not that egregious.
https://imgsli.com/Mjg2NjI3Which one is the best to you? >Hatsune Miku skateboarding in New York at night, neon, Watercolor style
>>101844379max_shift:1.15
>>101844370thanks>A painting by Beatrix Potter. The painting has wide brush strokes and visible paint daubs. The painting depicts Mrs Rabbit baking cookies in her kitchen.
any news on inpainting? that's what i'm waiting for to get on the flux train. gotta get rid of the butt chins somehow.
fashion snap, 80s, 90s>>101844342whats optimal?I copied & pasted animagine's recommendation
>>101844342>no, there has been studies on it earlier than that.but they gave shit results before, now it's a viable method because flux is a DiT model (transformers models are known to be resiliant to quantization) and it's a big model (the bigger it is, the more resiliant it is to quantization, we can see that on LLMs for example)
>>101844419>gotta get rid of the butt chins somehow.you can't add that on negative prompt?
>>101844379I like the line work and color palette for Miku on 1. It also has the better hands. But the ringing artifacts are more noticeable1.15 color palette overall matches 'neon' prompt better
>>101844352>I was just irked by the claim that it's more accurate than FP8.Same anon, same, they really didn't need to oversell their product, the simple fact that nf4 is close to fp8 while being 2 times lighter is already a really strong argument, especially for the VRAMlets
>>101844429i think that's more likely to remove BUTTS than fix the chins.
>>101844403
Some of my gens were really fucking up on latent upscaling -> add noise -> resample workflow on Flux - I think because of the depth of field.https://openmodeldb.info/models/4x-NomosUniDAT-bokeh-jpgThis upscaler worked best for me and didn't generate any artifacts.
>>101844450desu, the "boring photoshop 2015 circa" prompt can help aswell, you'll get more natural girls instead of fucking but chin Angelina Jolie's ones
seems to kinda know bill cosby
>>101844490but what if i want a normal girl in a different setting? normally i'd use [prompt|editing] but im not sure if thats supported yet. thats my bread and butter for SD.
still getting used to boomer prompting after years of learning how to get the most out of tag promptingtrying this as a negative (yes, cfg is above 1):>This image is incredibly ugly and disgusting, and also poorly drawn.seems to work lol
>>101844532kek, care to show the result anon? :v
>>101844260sdxl, not flux
>>101844450I wouldn't really use flux for anything except memes and landscapes desu.
>>101844260>>101844558thought nf4 required 30x+ nvidia?
>>101844614his model says Animagine so likely he's just retarded
>>101844371I read Euler a, nf4, and 20 steps.but not the model. lmao
>>101844423very true, anon.
>>101844672kekneeds some lawry's seasoning salt too
>>101844397>>101844435Thanks for your feedback, I'm trying to find a sweet spot but it really depends on the prompt, max_shift is a hard motherfucker to understandhttps://files.catbox.moe/91ea1q.jpg
>>101844614dunno but "nf4" works with less ram and less vram than "auto"idk why
https://www.reddit.com/r/StableDiffusion/comments/1eplvi8/comment/lhp0rfc/?utm_source=share&utm_medium=web2x&context=3>fp16 is a half precision distillation of the true model. it's weights might not be in an optimal position yet. NF4 seems to have disturbed interesting latents out of them by normalising groups of weights into 4 high precision numbers. it'll really be interesting if BFL puts something out in this line themselves. They potentially have a 32bit version of dev that they could quantize too.>Flux Dev really has no footing to call itself the "true flux". It's a development model intended for experimentation.That's really interesting, do you think they trained their model on 32bit though?
>>101844730yeah I don't know if you can assume they trained in fp32. I've heard some LLM researchers are even training in fp8 now
There is do little documentation on thingswhat the fuck does max_shift do?
Flux really blows everything out the water but it can't do nsfw. Is anyone finetuning an NSFW checkpoint?
>>101844770we didn't have good SDXL models for months, so it'll likely be a bit before anything real interesting hits the scene
>>101844730>>101844739no. it's all tf32, bf16, nf4, fp4 + AMP. varies wildly depending on arch, arch layers, training data, various hyperparams, and available compute.
>>101844746It's some shit invented by the SAI dev during the release of SD3https://huggingface.co/blog/sd3>To support inference with SD3, we have introduced a new scheduler (FlowMatchEulerDiscreteScheduler) with a rectified flow-matching formulation and Euler method steps. It also implements resolution-dependent shifting of the timestep schedule via a shift parameter. Increasing the shift value handles noise scaling better for higher resolutions. It is recommended to use shift=3.0 for the 2B model.https://reddit.com/r/StableDiffusion/comments/1di45dg/comparison_of_shift_values_in_stable_diffusion_3/>The Shift affects different aspects of the image, with larger values affecting the overall structure of the image and smaller values affecting detail. I figured a smaller value may work better for the upscaler, and it really does. Also allows you to up the noise significantly whilst maintaining composition. Shift of 0.5 and start step of 2 of 12 (equivalent to a denoise of 0.83) is giving me pleasing results.
>>101844781Got it. Thanks.
>>101842860>>101844462I'm going mental, but I can't upscale/resample this image (in any way) without noise artifacts. Image upscale, Latent upscale, Ultimate SD Upscale node. Nothing seems to deliver results that aren't just borked, even without injecting additional noise. The problem stops if I go for ridiculously high de-noise levels, but obviously that isn't feasible for a upscale detail pass.Other images are fine with this workflow. Could it be possible that Flux has some potentially tampered with training data that got into this gen? I know there was that whole 'anti ai hidden watermark thing' at one point.Here's a box for research purposes, but I assume the same thing will happen on any img2img workflow (ignore the prompt autism, it's chatgpt):https://files.catbox.moe/esvufo.pngAppreciate the help anons.
I would be curious to know what the flux prompt for this would be.
>>101844823Upscale example. You see the very aggressive noise and banding on the image? Happens with every method I've tried on this specific gen.
>>101844826looks like a pony/animagine gen, doubt its flux. Flux hates making nice heavy cleavage
>>101844823>I'm going mental, but I can't upscale/resample this image (in any way) without noise artifacts. Image upscale, Latent upscale, Ultimate SD Upscale node. Nothing seems to deliver results that aren't just borked, even without injecting additional noise. The problem stops if I go for ridiculously high de-noise levels, but obviously that isn't feasible for a upscale detail pass.can't you apply an AdaptiveGuidance threshold through it?
>>101844844It's pony. Flux doesn't really do large breasted anime girls afaik. I mean, maybe it could if you wrote 500 word boomer prompt and 500 word inverse negative prompt lol.
what's comfy's equivalent to the XYZ plot?like if I want to test different step values? 20,24,26,28what would be the best way to achieve this?
>>101844826Getting ChatGPT to 'describe this image for an ai generator prompt' has given me consistently good results for recreating things in Flux. Boomer prompt incoming.>This image features an anime-style character with cat-like features. The character is depicted with long, dark hair and cat ears, which are black on the outside and pink on the inside. She has a dark complexion and is smiling brightly, showing off her sharp teeth. Her eyes are large and amber-colored, giving a warm and cheerful expression. She is wearing a white bikini top that accentuates her figure, and there's a fluffy, light-colored garment draped around her shoulders. The background appears to be softly lit, possibly by sunlight coming through a window, giving the scene a warm and inviting atmosphere.>The art style is highly detailed and polished, with a focus on realistic lighting and shading that adds depth to the character and environment. The use of soft gradients and warm colors enhances the overall mood of the image. The character's expression and pose are dynamic, capturing a lively and cheerful moment. The style combines elements of traditional anime with a more modern, semi-realistic approach, especially noticeable in the rendering of skin tones, hair, and fabric textures.>The image is framed in a close-up shot, focusing primarily on the upper body and face of the character. The camera angle is slightly upward, enhancing the character's confident and playful expression. The proportions are exaggerated in typical anime fashion, with large, expressive eyes and an emphasis on the character's chest, which is prominently displayed in the composition. The close framing and intimate angle create a sense of closeness and immediacy, drawing attention to the character's smile and vibrant personality.
>>101844894you mean this?
>>101844900
>>101844900Not bad at all! Missing some style stuff but it has the basic idea down at least.Also, lmao:> The proportions are exaggerated in typical anime fashion, with large, expressive eyes and an emphasis on the character's chest,
https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/981>Using this option, you can even try SDXL in nf4 and see what will happen - in my case SDXL now really works like SD1.5 fast and images are spilling out!
>>101844953>>101844944Forge is back? neat.
>>101844881That's already in the workflow, 0.994 threshold - cfg 5.0. Tried with it removed as well and it had the same issue.>>101844900Also just as a sanity check so I know that my upscale method does actually work outside of that one fucked image lol>>101844942Unironically you've got to be really autistic with Flux prompts. For realistic women I end up saying shit like "she has thick thighs because she does lots of squats", because the prompt wont listen to you otherwise. It really doesn't like exaggerated proportions.
>>101844915no idea cause I'm new to comfyin A1111 you would use XYZplot for that, but maybe there's a better node for that in comfyI just want to try out different steps value, how would go about it?
Hi there sorry for the spoonfeed me post but I haven't kept up with AI since the disastrous SD3 launch so I will ask:Did anyone manage to unfuck SD3 or does it still give deformed bitches on grass?I am also seeing new base models on civit so I am assuming that non-SD diffusion scene is blooming.Which one of those are worth trying? At least somewhat comparable quality to SDXL/Pony hopefully.Also do they work on automatic1111 or do I have to learn another UI?(Please don't be comfy at least)>Why not try them yourselfCucked internet with quota so yeah I am limited in what I can download. Wanted to ask recs.
>>101844961Install "ComfyUI-nodes-hnmr" on comfy Browser, and add those 2 nodes and you're good to go
>>101844991based, thanks fren
I have a question, when using a lora, when there's multiple trigger words, does that mean we have to use them all or it can be used separately?
>>101844318Bring back Emma posting
>>101844998I'm blind now :(
but it works
>>101844960Yeah... LLMs have to be pretty indirect these days, and prefer to use terms like "form" or "figure" or "outline" instead of body. And they love to tie-in physical traits to other characteristics ("her posture reflects her earnest and upright character..."). You can go really over the top and be like "Her outfit highlights her generous assets, which are ample enough to make a man of the cloth begin to reconsider his vows." T5 is probably pretty good at parsing euphemisms like these.
>>101844900flux prompting is so weird. really does require having slight autismwill we ever get a danbooru type prompting style model...?
>>101844960>That's already in the workflow, 0.994 threshold - cfg 5.0. Tried with it removed as well and it had the same issue.try to decrease the threshold, the less the value is, the more it's gonna be "clean"
>>101845042One anon mentioned that he feeds his idea to a chatbot like claude and has it spit out a prompt format for him.
>Go to /ldg/ >Sneer at sub 24gb loser vramlets for holding the space hostage with their shitty hardware>Go to /lmg/ complain that insane multi 24gb GPU setups are the norm there and demand smaller modelsAnyone else /doublestandards/?
>>101844988There's a new player in town called Flux. SD3 is dead but this new model is basically what it should have been.Follows text really well and already as a base model kicks everyone's ass.Fairly censored though, you can get tits out of it but it doesn't want to do sexual stuff.Now we just wait for people to work their autism on it and we'll have something locally resembling DALL-EWorks on Comfy and Forge.
Can you only load NF4 via a new loader in comfyui? Does it work with a separate dualclip loader for fp16 t5xxl?
>>101844836i noticed the same thing in pedoanon's gens. they were full of little squares.
>>101845106It's not really any more or less censored than SDXL is. I think we should hold off on celebrating too soon though. Even if we can finetune it and make LoRAs for it, there's evidence these training efforts fuck the model is quite a few ways.
>>101845131Those are watermarks. A post on reddit said they can be decoded and reveal things like GPU make and other things. It's possible pedo joe doxxed himself.
>>101845119>Can you only load NF4 via a new loader in comfyui?yes you canhttps://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4>Does it work with a separate dualclip loader for fp16 t5xxl?yeah you can load a separate fp16 t5 with it, forge sucks in that regard
nf4 works on every chip as long as you hope
>>101844929australia in the eye reflections
can LoRAs be trained with NF4? If so, we may be back. If not, fuck off.
>>101845034LLM tagging was a mistake
>>101844960>>101845034Following-up...Positive: A high-quality, anime-style image of a young woman with a cheerful expression. She has brown skin and bright green eyes that are slightly narrowed as she smiles widely, showing her teeth. Her hair is short, straight, and black with bangs that frame her face. She has black cat ears with light pink inner fur that match her dark hair. Her face is slightly flushed, giving her a playful, teasing look. She wears a dark, simple tank top, revealing her shoulders, collarbones, and ample assets, which are generously proportioned. The background is a light, neutral color to emphasize the character.Negative: A low-quality, Western-style mosaic of an old man with morose expression. He has sickly skin and grey eyes as he keeps his lips pursed shut, nursing a sour expression. He wears a trench coat, concealing most of his physique which is okay because nobody wants to see what a sour old bastard looks like anyway.Guidance: 7, euler, schnell.
>>101845166at this point, everything should be trained into a bitnet setting (1.58bit) so that we would get giant models running for everyonehttps://arxiv.org/abs/2402.17764
>>101845183Please don't bring the bitnet meme to /lmg/ it's not real.
>>101845193it's not a meme, it worked well at 3.9b
>>101845042>will we ever get a danbooru type prompting style model...?i don't think this is really a good idea. you lose the ability to describe the position of things and relations between objects. the truth is that we need for the base dataset to be human annotated with precise description free of verbal diarrhea.
>>101845206The models are flexible enough that you can provide examples of both in the training data and get a mix of both worlds.
https://new.reddit.com/r/StableDiffusion/comments/1epugio/contrastive_loss_custom_loss_pushing_clip_aka_the/>Pushing CLIP aka 'the text encoder' to 91,4% ImageNet/ObjectNet accuracy. Wtf? there's a better clip_l in town, you should try it outhttps://huggingface.co/zer0int/CLIP-GmP-ViT-L-14
>>101845206Can't wait to see what the prompts look like if/when something similar to flux comes out for porn."Anon gently kisses Miku and sends shivers down her spine....LARGE BREASTS"
>>101844953>>101844944sweet, I'm going to test forget
any advice installing forge compared to A1111?
>>101845253>Help I cant press the one button install
>>101839583https://github.com/scraed/CharacteristicGuidanceWebUILooks like it's an alternative to DynamicThresholding, if it's better I'll provide some feedback
>>101845106Thanks anon, it sounds promising.This is the model I should install right?https://huggingface.co/lllyasviel/flux1-dev-bnb-nf4/blob/main/flux1-dev-bnb-nf4.safetensorsAt least for my 12gb card.
I don't know where else to ask so here goesHow do I run a local alphafold on my computer or pay for cloud computing to do it on a server? I'm just trying to get pdb files from fasta files via alphafold prediction
>>101845261that's not what I meanI mean if there's any relevant difference between forge and A1111
>>101845262if it really can remove the "white blur" dynamic thresholding used to have then it's a big fucking deal
>>101845262>>101845274https://github.com/redhottensors/ComfyUI-PredictionIs this its comfyUi equivalent?
>>101845263Yes that's the right model.
guys I'm demoralized
how the hell do I run nf4? i downloaded, put in my custom nodes folder but i cant find the actual node anywhere and not on the manager either
>>101845322did you update comfyUi?
40 seconds if i don't mess with the prompt, nf4 is pretty great. all we need now is loras on nf4
>>101845181My negative prompt is starting to crack me up:> A low-quality, Western-style etching of an old man with morose expression. He has sickly skin and grey eyes as he keeps his lips pursed shut, nursing a sour expression with one of his eyes closed. He wears a grey poncho, concealing much of his physique which is okay because nobody wants to see what a sour old bastard looks like anyway, although his chest is clearly concave and sunken inward, making his upper torso appear hollow. The image continues in an awkwardly angled direction, making the man's head look unnaturally large, while contorting his shoulders in a strange manner. The image is also watermarked and contains messages designed to chastise the user for wanting to generate such an image.I'm not sure if it's all getting through lol.>>101845237Flux girls are gonna feel the ministrations reverberate through every molecule in the room as the heat in their core coils tightly within them before surrendering to the tidal wave of pleasure that threatens to consume them ... a lot.
>>101845001In the case of multiple, it's usually optional to use any at all. Some models include minor variations of the same thing that use different trigger words, but unless they say that specifically in the description assume that any trigger words are optional and are only there as a recommendation to help you get closer
Hello, I need help. I found an iPhone 15 in a taxi this afternoon.
>>101845320Well, at least you are not ropeless :)
>>101845407Give it back
https://github.com/comfyanonymous/ComfyUI_TensorRTDoes this work with flux?
>>101845407give it back
I found it in a taxi i dont know
>>101845274 >>101845299 https://github.com/scraed/CharacteristicGuidanceWebUIthe fuck is "cond" "uncond" "fallback" "prediction", how I'm supposed to make that work, it's way easier on DynamicThreshold
>mfw trying nf4 flux dev in comfy on rtx 2080Holy fuck, the entire thing fits into my 8gb of vram with a separate dualclip loader AND I can use fp16 t5xxl. Forge gives me 9-10 sits and it chunks a huge load into ram, and that's with a shitty fp8 t5xxl
>>101845486Grats, real happy once again that we're all taking a quality hit to please the vramlets. Good for you.
>>101844770>>101844826>>101844893>>>/h/hdg>>>/e/edg>>>/b/degen>>>/d/ddg>>>/aco/sdg>>>/trash/sdg
>>101845449Unrelated but this post made me remember that tensorRT exist.Does anyone know roughly how much performance I am leaving on the table by not using it during image gen on a 3060?
>>101845392
What works is what's good.Be happy with what works.
>>101845449I tried to do it but couldn't figure it out. It kept failing while trying to build the engine.
>>101845549
It's insane to know how much techniques has been invented so far to prevent the burning of high CFG, I'm pretty sure that the current combo of (CFG + DynamicThresholding + AdaptiveGuidance) isn't the best one, I'm gonna test them out and see if I can improve flux even further
Can anyone share aworkflow that has...NF4, + Lora support + Negative prompt?Ive been playing around with several and havent foudna way to make all of these work
>>101845237i lol'd, but that's what i've been doing so far in pony/SD. simple boomer sentence followed by tags.
>>101845497You can load fp8/fp16 just like you normally do though.>>101845583Just grab a normal workflow with dynamic thresholding that's been posted here numerous times and replace a regular flux model loader with a NF4 one. Then connect a LoralLorder node to it, and then dynamic thresholding to your lora and you're good to go. If you need multiple loras just connect multiple lora nodes to each other.
>>101845583I would like this as well, but I don't think loras work with NF4 Flux yet. Hopefully soon.
>>101845576
>>101845581https://github.com/Extraltodeus/Skimmed_CFGkek, that one didn't change jack shit, I got the same exact output as the raw CFG 6
>>101845485https://github.com/redhottensors/ComfyUI-PredictionSomething like this? idk I'm not tech savvy
>>101845802I gave up on that shit, they said it's slow as fuck I don't see the point, DynamicThreshold already decrease the speed by 20%
>>101845803welcome back, no tods please.
>>101845698
>>101845846One more.
>>101845753That's funny... Perturbed Attention Guidance (PAG) seems to be doing on flux even though it's supposed to only work on unet models, I expected an error or some shit kek
>>101845817Yeah, did it exactly as on that pic and it doesn't even work, fuck that
yea prompting with flux is so autistic i really hate it. it's a really good model but it really feels like how i felt back when NAI model was leaked. i miss this wonder.
>>101845914PAG seems promising as fuck, I start to get working picture, maybe combined with DynamicThreshold would make the quality even better, need more testinghttps://github.com/pamparamm/sd-perturbed-attention
NVIDIA Genesis Evangelian
>>101845720catbox please?
https://www.reddit.com/r/StableDiffusion/comments/1epl454/flux_dev_lora_training_use_simpletuner_it_works/Reddit trained a working character LoRA on a single 4090 for Flux dev. Absolute madlads.
>>101845926I only ever used the regular perturbed attention guidance node, along with self-attention guidance and freeu. The seem to drastically improve the quality with the right settings, but I was using them on sdxl and I imagine it would need completely different settings for flux, if they even work.
>>101842860is this the non tranny version of sdg?
>>101846114>I imagine it would need completely different settings for flux, if they even work.I don't know how it doesn't even crash, it's supposed to only work on unet models, flux is a DiT, but hey, it works and is giving different results based on the scale so I take it lmao
>>101846124>>101846099Do you have a fucking brain tumor or something?
>>101846124>>101846099uh anon u alright there?
>>101846127Damn, I didn't even know it goes that fatr These were my settings for sdxl
>>101846144>>101846127forgot a pic
>>101846156you seem more knoledgable about that topic, maybe you shoud make some experiments on flux and see how far it can go
>>101846165Nah, I just saw some posts on civitai that suggested the settings, like https://civitai.com/articles/5761/tuning-with-freeuWith SAG I just fiddled around until I found something that looked generally good for realistic stuff and anime. Then again that was for SDXL so I doubt it matters much for Flux
>>101845449I couldn't even build the wheel for tensorrt-cu12 while installing lol. I found some answers on the net saying that it's impossible on windows and recommending to download some containers with prebuilt wheels, but I'm not going to spend hours learning docker and fiddling with unfamiliar soft just to save some seconds generating.
is there really no way to force ComfyUI to run the text encoders on the CPU to stop all this swapping between RAM and VRAM?
>>101846191it's possible yeahhttps://reddit.com/r/StableDiffusion/comments/1el79h3/flux_can_be_run_on_a_multigpu_configuration/instead of going for cuda:0 or cuda:1, you choose "cpu" for the text encoder
>>101846191easy >>101689729
>>101846184oh ok, thanks for the civitai, I'll see how that in detail aswell
>try FP16 T5 with old gens to see if there really is a difference>small aesthetic differences>text previously legible becomes fucked upwhat the hell, anons, you lied to me
>>101846156Ok I tried Self-Attention Guidance and I got an error, that one definitely needs a unet architecture to work
>>101846245>>101846156Same thing for FreeU, at least the advantage of a DiT model is that there's less choice so we can focus on a few things at the time kek
>>101846246What did she do?
>>101845926>PAG seems promising as fuck,SEG looks even better, nice
>>101846275killed and butchered five families
>>101846275Said the N word on the internet
>>101846263Business idea: Fresh ice cream masticated by robot girl
>>101846204Not him but thanks, it solved my current problem of t5xxl fp16 taking ridiculously long time between gens with nf4 on comfy. Now my overall time for nf4 gens is finally lower than with fp8.
https://huggingface.co/alvdansen/frosting_lane_fluxNgl this is pretty neat, is a manga style lora all Flux needs?
oh no its this fuckin guy again
Even when Flux doesn't know a character, it can get pretty close with a good description
>>101845581Holy fuck, I found a node "SEGAttention" that gives good output for CFG 6 without the need of DynamicThresholdinghttps://github.com/logtd/ComfyUI-SEGAttention
>>101846616The pics on github look good, your idk. Try a complex prompt with a very heavy style that base cfg can't handle.
>>101846643i used the default values of 3 and 10, fortunately unlike dynamicThresholding it's only 2 parameters so it'll be easy to find the right combinaison
>panties>T5: okay>striped panties>T5: you mean shorts?every time
>>101846643>Try a complex prompt with a very heavy style that base cfg can't handle.do you have something in mind?
>>101846675Sir this is the technology board
>>101846682>An illustration inspired by the works of Jean-Baptiste-Siméon Chardin. The scene depicts a 18th-century classroom with Hatsune Miku as a teacher at the center, guiding a group of attentive young students. Miku is dressed in her classic outfit altered to fit to the setting. The classroom is filled with wooden desks, chalkboards, and books, all rendered with Chardin's characteristic focus on realistic textures and warm, muted colors. The words "How to prooompt" are clearly written on the chalkboard. There is a speech bubble coming from Miku's mouth with the text "Don't be a promptlet". The lighting is soft and natural, creating an atmosphere of calm and scholarly dedication.
what the hell, this isn't shipaman
>>101846548Not a bad lora
>>101846712https://imgsli.com/Mjg2NjkxDynamicThreshold wins but SEGAttention still has the default values so it can be improved I guess, more testing needs to be donehttps://imgsli.com/Mjg2Njkz
>>101846774Seems like that you can't put both SEGAattention and DynamicThreshold, only SEGA is activated when it's the case
>>101846548Turns out flax is not only trainable but apparently quite easily trainable (sans hardware)Fine tunes when
>>101846836When quant 4 LoRA training?
>>101844787thanks, anon!
>>101846836I'm retard, does the fact that loras are easily trainable mean that the model itself must also be easily trainable?Also is it possible to run multiple loras at once?
https://civitai.com/models/640156?modelVersionId=715962Booba and Cocka
>>101846836Efficient single 3090 training when? QLoRA is a thing after all
>>101846887I'd happily rent some gpu time to train some loras but I want to wait for the community to do all the hard work refining the process first.
>>101846881That was fast... it took us months to make good NFSW anatomy on SDXL
If I have a 8gb and a 12gb GPU is it still worth offloading the text encoder to one and the rest of the model to the other, or should I just use nu-forge with only the 12gb?
>>101846881doesn't work with flux1-dev-bnb-nf4.safetensors :((
>>101846881Nipples look pretty good, but genitals definitely still need some work
>>101846887Dude trained one on a 4090 so who knows
>>101846712>>101846616false flag, SEGAattention gives the same output whether we're on cfg = 6 or cfg = 1, but on cfg = 1 it can do negative prompt though, and it's not 3x slower like on PerpNegAdaptiveGuider, it's the regular speed you get on cfg > 1https://imgsli.com/Mjg2Njk4
>>101846933You redraw the nipples at least using auto Booba masks and running a pony model over the output>>101846944Yeah still barbie vag but at least the underwear comes off, that's more than what base can do.Guys, I have a feeling Flux is the new community default model. In a week we have nips and cocks and vramlets running it. Pony is wasting his time and money with AuraFlow
>>101846919for some reason the whole thing fits in 12GB for me now, i used to offload to regular RAM back when flux came out but not anymore, i think they updated something. note: i'm retarded so if i'm wrong please tell me why
https://civitai.com/models/639737/emma-watson-lora?modelVersionId=715459
>>101846956That's neat, when someone figures out how to do it and posts their code we'll be eating good
>>101847026He did it via Ubuntu wsl, apparently had to use his cpu for video output and it was at 100% use
>>101847005huh nice if it fits within 12gb like that
>>101842860People still remembered to include AuraFlow? It got mogged so hard by Flux.1 I completely forgot about its existence.
>>101847024so it begins
New day, new bread...>>101847080>>101847080>>101847080
You CAN'T run fluxok you can run it but only on 4090sOk you can run it on high end GPUsOK you can run it on mid range GPUs but you can't train itOK you can train it but only on data centresok you can train it on consumer GPUs but ONLY loras!>WE ARE HEREok you can fine tune it but it won't beat SD!Where the SAI shill now? Suspiciously quiet, pretending he didn't say any of the retarded shit he said. I want to see more cope.