Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101741569>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>Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
>>101744342How is it possible that after so much time and samplers Euler A is still the king? When are they going to come up with something better, bros?
>>101744381kek
>>101744412>06.08.1024Trump's been around a while
Tried heunpp2 and it looked way worse than euler. Guess it depends on the gen but not really worth it to me if it's 3x slower and not consistently better
>>101744426for me the result was better, but not by much, and it took 3x longer. It's not a big of enough of a difference for the performance hit for me.
>>101744477o snap
>>101744484Why does right have dwarf proportions, bros?
trump at the porn store lookin for a masturbator
Trump deciding whether to buy the gamer girl pee or the epstein's select cunny juice
>>101744425
>>101744536He went to the wrong sex shop
>>101744388it's funny cause in the technical details of how these things work, euler's method is a super basic and simple approach yet I've never found the fancier ones to be a clear improvement
>>101744562Oh shit they brought in the double barrel tank
>>101744388>How is it possible that after so much time and samplers Euler A is still the king?Only euler works on flux, not euler A
Generate Miku or else
>>101744555can you go for trump going into the cunnybot store? kek
>>101744641it's funny how flux doesn't know with great accuracy much of the celebrities, but can do the most controvertial and AI-hating one easily kek
You guys think quality negatives will be of any use with flux? I'm trying out "The anatomy is warped and unnatural, appearing mutated with limbs bending and twisting in impossible ways."
>>101744686schizo negative bros is our time to shine?
I don't get the hype. It just feels like SDXL finetuned on the stuff normalfags liked about dalle like the 3d renders, photos with overblown DoF, cctv cams, etc. It's pretty awful at styles outside of that. It's good for a base model I guess? The only really impressive thing about it is the text.
>>101744700Yeah idk, flux seems to genuinely understand prompts at a level beyond just tags unlike previous models so maybe it can have an impact
>>101744711>I don't get the hype. It just feels like SDXL finetuned on the stuff normalfags likedSDXL can't do something as complex as this anonhttps://www.youtube.com/watch?v=N00r4U2--eM
>>101744388I've been using DPM++ 2S a with considerable success
>>101744711It's the coherence, nigga.
>>101744711Nah, it's a genuinely smarter architecture and it shows. SDXL still felt like all it could do was understand booru tags, flux is different
>>101744711it understands promps so well, the image quality is top notch, the text is probably the best of them all (local and API considered), for a base model it completely destroys everything we had so far and you don't see the hype? lol
>>101744702YAY HER HANDS ARE DISFIGUREDWhat's your prompt?
>>101744711>Type men screaming in horror over stock market crash>get men screaming in horror over stock market crash It has great meme potential even if the specific artistic styles are weaker then some other models. Everyone is waiting for the finetunes that are (hopefully) coming. It's just neat that I type what I want and get what I want. +text of course.
>>101744744>thigh gapI'm
>>101744773dev said finetuning flux was impossible
>>101744781I'M WEARING THIS SHIRT TO CHURCH
>>101744773>It's just neat that I type what I want and get what I want.yeah, it's great, but not perfect though >>101744136
>>101744795Check back in ten years.
>>101744605picrel, but I would love to do it in >>101744674 style. What's your prompt?
>>101744782>dev said finetuning flux was impossibleSo impossible it's already been donehttps://xcancel.com/ostrisai/status/1820462674230059328
>>101744786#badfeet>>101744795I'm working on it.
>just realized I spend the last 14 hours so prompting
>>101744804lol
>>101744804https://files.catbox.moe/n11v3g.png
Can anyone spare an upscaling Flux workflow?
>>101744814>mfw I have been up all night proompting
>>101744804This model is really amazing, can't believe we're running this locally, feelsgoodman.
>>101744875yeah.. even when it fucks up its still 1000x better than SD
>>101744866its too much fun
>>10174482720 steps on schnell?Schnell is a turbo model, you can gen on 4 steps.
>>101744890o.. i didn't know
>>101744866Leave your pc queued up :) come back the next day and reap your sweet gpu dividends.
>>101744711>ArtifactsArtifacts still exist, but they are less atrocious and are hidden much better relative to SDXL>HandsHands look like hands. it gives extra digits occasionally still but the tentacles problem is solved>Adherence to promptAdherence to prompt has never been seen at this level before (complexity + coherence), even surpassing dalle-3 (except maybe in terms of style variety)>Text"Text" is underselling it. full sentences of fully legible text in 3-4 tries using only text prompt. This is unprecedented>Open sourceNo one can take FLUX from me, I have it and it's mine and it runs on my consumer-grade hardwareThis also opens the door to finetunes and loras (despite what doomers say)>Small company competing with OpenAI"Open"AI has much more capital and has only managed a comparable product. When BFL starts raking in revenue, there is the potential to far surpass oai
>>101744890What's the tldr on shnell?
>>101744907>"Text" is underselling it. full sentences of fully legible text in 3-4 tries using only text prompt. This is unprecedentedthat's ironic because the main selling point of SD3 was the text and all it could do was some ugly comic-sans photoshoped text, flux mogs SAI so hard
>>101744907We definitely need negative prompts.
>>101744918kek.. i didn't know sd3 did text.. sd3 sucks balls
>>101744916it's the SDXL Turbo of flux, it's fast but the quality is obviously worse
>>101744927so dev is better?
>>101744925>We definitely need negative prompts.you have 2 optionshttps://reddit.com/r/StableDiffusion/comments/1el3tnq/want_to_use_negative_prompts_with_cfg_1_on_flux/https://reddit.com/r/StableDiffusion/comments/1ekgiw6/heres_a_hack_to_make_flux_better_at_prompt/
>>101744859https://files.catbox.moe/4p3hol.pngConnect these 2 here to turn on the upscaler.
>>101744938of course, that's why they didn't put an apache 2.0 licence on dev, they don't want us to make money out of that superior model
>>101744938Dev is better Schnell is faster, both kick the ever loving shit out of SD3
>>101744925>The way I've always done it is the way it NEEDS to be done. I am incapable of adapting to change
>>101744954You can make money off the gens, you just can't train and sell a competing model using it as a base.
https://www.reddit.com/r/StableDiffusion/comments/1el79h3/comment/lgpz422/?utm_source=share&utm_medium=web2x&context=3>Just a heads up: there’s an “all in one” FP16 model on civit now that has everything baked in. (CLIP and VAE). It uses about 16GB of VRAM. You load it over the normal load checkpoint node. Leaves you plenty of VRAM to use your system besides.What? Chat is this true?
>>101744927Will shnell gen the exact same image for the same seed, cfg, and guidance?
>>101744966I have generated about 5k images since flux dropped and I have not once felt the need to have a negative prompt.
>>101744972>You can make money off the gensI don't think you can, the licence dev has seems worse than the "old" SD3 licence, the one that pony bitched about
>>101744963SDXL came out and was hardly worth bothering with over 1.5 until it was finetuned by the community into something good much later. Then SD3 comes out and I don't even bother trying to set it up, and I'm pretty sure I didn't miss anything. Sad!They need to release a SD4 that BTFO's flux or it's officially over.
>>101744983unless you're using xformers, image models are always deterministic yeah
>>101744998>They need to release a SD4 that BTFO's flux or it's officially over.they're too busy seething and finding excuses, it's overhttps://xcancel.com/Konan92_AI/status/1820518655450562588#m
>>101745000>unless you're using xformers, image models are always deterministic yeahhaven't heard that in quite a while, didn't the nondeterminism of xformers get fixed way back?
>>101744978Does he mean this?https://civitai.com/models/623224/flux1s-16gb?modelVersionId=696732Because it's FP8 not FP16
>>101744966Fine, let's see you generate a woman with a deformed hand, or a stump arm (amputation). Bonus points if there is a visible disfigurement, but apparent is also impressive.Prosthetics kind of gen, but are too good looking, so far.
>>101744989>the one that pony bitched aboutAt first I was intrigued about this guy, now I hate his guts. If he won't do anything with his dataset, it should go to someone who will.
>>101745036>Conversion of Unet to Checkpoint including T5 fp8, Clip L and VAE, which gives a model of 16GB.Yikes, fp8 t5xxl is a really bad idea, this text encoder should never be used as the quantized version, it fucking sucks at fp8
>>101744987But, you can't generate a woman with an amputated arm, or a deformity of the hand/arm, can you?>>101744942thanks
>>101745047>At first I was intrigued about this guy, now I hate his guts.Same, at first I thought he was the first non mentally ill ponyfag, but he lost all my respect when he cucked the artist tags on v6, fuck that bitch
>>101745054yeah, and that one only requires RAM which is cheap and easy to upgrade unlike VRAM, people should absolutely use the fp16 if at all possible
>>101744711>download model>1girl, spread anus, giant stinky asshole>no results>uninstall>"I don't get the hype"
>>101745075kekd
>>101744907>When BFL starts raking in revenueIgnorant question here. How exactly do they do that? Just by hosting image gen capabilities like NAI? I know they have a pro version but it's hard for me to tell if it's actually better or even worth paying for.
>>101745046yeah, flux can do thinks he knows great, but he also doesn't know a lot of things, but that also mean that it's just a matter of finetunes, you'll know that flux will nail those news concept we'll be putting on its ass, and that feels good to know that
>>101744989You didn't read the licence. You just clicked "I agree, I am a retard" didn't you.>Outputs. We claim no ownership rights in and to the Outputs. You are solely responsible for the Outputs you generate and their subsequent uses in accordance with this License. You may use Output for any purpose (including for commercial purposes), except as expressly prohibited herein. You may not use the Output to train, fine-tune or distill a model that is competitive with the FLUX.1 [dev] Model.
>>101745099>know they have a pro version but it's hard for me to tell if it's actually better or even worth paying for.not a lot of people have a 24gb vram card or are willing to wait long minutes for a single gen, API market will always be lucrative
>>101745099>How exactly do they do that?The same way every tech company does, rake in speculative capital and when the investors come back to get their profits the store is vanished as if it were never there to begin with.
Someone explain how we can exploit this
>>101745104>You may not use the Output to train, fine-tune or distill a model that is competitive with the FLUX.1 [dev] Model.:^) Can't, or you wish we wouldn't?
>>101745121>that is competitive with the FLUX.1 [dev] ModelYou can't sell it, or access to it.
>>101745104>including for commercial purposes>commercial purposesthat means you can't make money out of your expensive training, like making a kickstarter or some donations right? if that's a yes then it's DOA
>>101745133Can't is a word which I think you do not understand.They wish you wouldn't. :^)
>>101745110noice, what prompt did you use anon?
>anons don't understand what competitive means
>>101745119
>>101745134>You may use outputs>OutputsOutputs are the generated images.
>>101745099While dev and schnell are opensource, BFL offers an API service to gen them on their servers for a feealso there is also a third "pro" model that is closed source, only available through apireleasing the open-source models could be interpreted as an advanced form of advertising
>>101745169but can someone make money out of his training via donations? that's my question and you didn't answer it
>>101745159Surely this is more effective.
>>101745104Does the license also forbid others from hosting the model and charging for gens on a different website? (Like civit)
>>101745119Why too much My Little Pony and not enough anime babes. What the fuck was Black Forest thinking?
>>101745046Tried peg legs,pirate hooks,disembodied limbs,invisible this and that and nothing seems to have worked so far.
>>101745144why do people keep asking this? type the things you see into the prompt field. its really that easy
>>101745234It's not fair. The model knows all the ponies but not characters like Tifa.
>>101742866Thanks fren
>>101745185https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.mdRead nigga. Training a model for community use is permissible as long as it remains within the scope of non-commercial use. You cannot use the FLUX.1 [dev] Model or its outputs to train a model for commercial use without obtaining a separate commercial license from the company.The definition of "Non-Commercial Purpose" includes any use where you do not receive direct or indirect payment from the use of the model or its output. Collecting donations falls into a gray area, as it could be interpreted as indirect payment. However, if the donations are used solely to support the non-commercial project without any commercial intent or benefit, it might be acceptable.You can train a model for non-commercial community use but there's no information about collecting donations.
>>101745277<can
>>101745277This is probably why ponyfag said no to a fluxpony.
>>101745265>type the things you see into the prompt field. its really that easyit doesn't always work, for example that image >>101744875I first tried with a simple prompt which was:>Hatsune miku laughing hard, pointing her finger towards her computer screen, she says on a speech bubble "Trump really went into a store!it made a profile picture of Miku pointing her fingers at herself, I then decided to make it more descriptive with Claude 3.5 Sonnet>A digital illustration of Hatsune Miku. She's sitting at a desk, leaning back in her chair and laughing heartily. Her eyes are squeezed shut with mirth, and she's holding her stomach with one hand while pointing at her computer screen with the other. The computer screen is visible, showing a news website. Above Miku's head is a large, cartoon-style speech bubble. Inside the bubble are the words "Trump really went into a cunnybot store!" in a playful font. The overall scene has a bright, anime-inspired art style with vibrant colors.and then it worked, flux seems to prefer verbose prompt
>>101745295No, Pony is saying no because he already invested shit loads of compute into another model.
>>101745277I don't give a fuck. I'm going to train Flux and I'm going to make money out of it. If Nigga Forst has an issue with it, they can sue me.
>>101745277This is a good thing. I don't want those fucks at openai, micropenis, or any others to profit off this gift from The Lord.
>>101745295Why does the guy with 8 H100s in his garage care about money?
>>101745277>Read nigga. Training a model for community use is permissible as long as it remains within the scope of non-commercial use.so it's fucking DOA, no one will be willing to spend thousands of dollars in finetune and not having any money in return
>>101745198You gave me the idea
>>101745314The only way we benefit from a private finetune is if it leaks and there's little chance that will happen again.
>>101745299Yeah I usually try to give verbose and kinda flowery prompts like I'm writing a novel or something. Seems to actually work better. I wonder what it is about the training process that makes it respond better to those descriptions.
>>101745314interval-based cyber begging
>>101745302Comfiest double wide I've shopped for.
>>101745314Wrong. People do it for free all the fucking time. Welcome to the open community, where everything is free and capitalism doesn't matter.
>>101744711just a bunch of newfags from /lmg/ really
>>101745352>Wrong. People do it for free all the fucking time.not at the scale of what flux is asking, that's a 12b model, it asks for a LOT https://github.com/bghira/SimpleTuner>Flux.1 [dev, schnell]>A100-40G (LoRA, rank-16 or lower)>A100-80G (LoRA, up to rank-256)>3x A100-80G (Full tuning, DeepSpeed ZeRO 1)>1x A100-80G (Full tuning, DeepSpeed ZeRO 3)
>>101745328Word on the street is that a VLLM (think GPT-4 with image input) was used to create descriptions for the images in the (apparently well-curated) dataset. Which explains a lot but until I see a source on that (or the documentation please (what's the token limit??)) it's just speculation
>>101745359no wonder we were inundated with miku posts
>>101745354that niggeress is also named Adolf Hitler kek
>>101745359An A100 costs $1/hour to rent.
>>101745389if you want to finetune the model to the point it knows what porn, celebrities and anime characters are, it will probably cost tens of thousands of dollars
>>101745396>it will probably cost tens of thousands of dollarsprobably way more, pony said he spent 30k dollars on v6, and that was only for SDXL, a 3.5b model, flux is almost 4 times bigger
>>101745405I don't think I've seen anyone else do flux anime like this.
Why is this general full of dumb as rocks doomers who get btfo at every dumb ass thing they say yet still continue?>You can't run Flux on consumer hardware!>Ok you can run it on consumer hardware but only on a 4090!>OK you can run it on cards other than a 4090 but you can't train it!">OK you can train it but it's too expensive!>OK it's not too expensive but nobody will do it!YOU ARE HERE>OK people will do it but not for free!>OK people will do it for free but they won't share it!>OK people will share it but, but, but...
Yo
>>101743143
>>101745405>yuria man of culture I see
>>101745396And there's no guarantee that the training will work and not break so many of the features that make the model good in the first place. I don't see training for this model going beyond a few choice LoRAs and controlnets desu.
>>101745389>A100>HF starts at $4/hr>Amazon starts at $4/hr>Google starts at $6/hrwhat Chinese scam are you trying to shill?
>>101745421I'm committed to SD no matter what. I made that pledge long ago. I don't care how much SAI may shit the ball. I'm going to support them to the end. If that means I have to lay and FOMO other models, I'll gladly do it.
>>101745430>there's no guarantee that the training will work and not break so many of the features that make the model good in the first place.desu there should be zero LLM captioning during the finetuning, especially if you want to simply add more trivia into the model, we have no idea what they did to make the model so good at prompt understanding, so making it simple would be the key to get improvements yeah
>>101745407It's getting hard to tell AI from real
>>101745307>they can sue me.Knock knock Anon>>101745421Funny thing is all those BTFOs are in the space of less than a week. I wonder what the new cope will be.
>>101745451you're posting in the wrong thread
>>101745451excuse me?? are you in a cult or something?
>>101745443The first link on Google that had pricing listed, which was the 2nd link on Google for the search "rent a100"
>>101745421When things go right it's a pleasant surprise instead of constant disappointment. I'm not a doomposter though.
>>101745405that's kinda incredible, for that one there's no way you can tell if it's an AI and not a random screenshot from an anime
>>101745451Hey buddy you got into the wrong door, the >>101745393 thread is 2 boards down
>>101745480These also look real >>101745312Flux does things that SDXL could only do with loras, and better
>>101745480The subtitles are misspelled. It really doesn't like the word "are"
>>101745443>>101745465Clicking some more links on the front page of Google:Hyperstack - $1.32/hourRunpod - $1.19/hrImmers - $2.35/hrPuzl - $1.60/hr
>>101745312>>101745505>"Blacks"even hitler is more respectful than us when we call the niggers kek
>>101745522So it's like what, 5 bucks to train a LoRA with a dataset of 100 or so images? That's not awful. So long as LoRA training actually works. I'd like to see more examples before I toss money at it.
>>101745516Leave the text out and just edit in your own subtitles.
Should I treat Flux.1 as if it just does CLIP encoding of text, or does it do more text processing?
https://www.reddit.com/r/StableDiffusion/comments/1el79h3/comment/lgq3riz/?utm_source=share&utm_medium=web2x&context=3>I want to test the multi GPU setup without comfy - like an ordinary Python program using diffusers.Dare I say based?
>>101745530Yeah same, but 1 trained 3 days after a new model comes out is impressive.Took him a day I think, I am sure efficiencies can be worked in to cut that time way down.
>>101745423Cool gen
>>101745527He had a good heart. Cared for animals, wanted to preserve the innocence of children and keep the European society traditional, safe, and stable.
>>101745540I think we'll see over the coming days the Vram requirements probably halve again when new workarounds are figured out. Best case scenario is 24gb cards can at least train something at a low rank locally. And that's probably all I'd really want for this model, just some loras to push it further in one direction.
>>101745558>I think we'll see over the coming days the Vram requirements probably halve again when new workarounds are figured out.that's not possible, the size of the model is what it is, everything has to go into the VRAM
>>101745573Even at FP8? Is it not feasible to train LoRA at that size?
>>101745463If being principled means I'm in a cult, then so be it. When everyone started turning on SAI, I deiced I was going to remain loyal to them. They gave us this. I'm not betraying them now they are on the floor.
>>101745573This model seems to use 40GB of memory.Why can't Nvidia just double the amount of VRAM on the 5090?
>>101745585>Even at FP8? Is it not feasible to train LoRA at that size?I think it's at fp8 by default on SimpleTuner?
>>101745532But that's not fun.
>>101745602Damn, even at FP8 its still not possible on a consumer card? I swear I saw somewhere yesterday that 24GB was possible at rank 16
>>101745594>This model seems to use 40GB of memory.the fp16 flux model asks for 22-23gb, and the fp16 t5xxl asks for 9gb, so yeah, it's into the 30gb range>Why can't Nvidia just double the amount of VRAM on the 5090?because they're selling 48gb vram cards 10 times the price of a 3090, there's no way they'll give us more vram, that would ruin their business plan
>>101745591>When everyone started turning on SAI, I deiced I was going to remain loyal to them.why? they aren't some god that can do no wrong, I respect their contribution of the past, but that doesn't mean that I should turn a blind eye about their shady tactics they're doing now, that's not how it works
>>101745615Next question is will it be possible to NVLink 2 3090's together to train it?
>>101745623hes trolling very badly, anon
>>101745615example.>>101745302
>>101745613>I swear I saw somewhere yesterday that 24GB was possible at rank 16here? >>101745369
>>101745615I'm really hoping Intel decides to save their business by stapling 48gb of vram to one of their cards and selling it for $2k. It doesn't even have to be good, it could have the performance of a 3060 for all I care, just shove a shitton of vram in there and it'll sell like hotcakes and people will make the software stack work for it.
>>101745631>>101745639AMD is a meme, Nvdia is fucking all in the ass because of Cuda, that shit is reponsible of their monopoly, and AMD's "cuda" is really really bad, no one will work with that shit
>>101745634Nah it was a github thread. I can't find it.
>>101745591SAI never gave you shit, not really, you know that right? SD was leaked by Runway and SAI decided to try and build a business out of that. Literally everything they have done since 1.5 has been dog shit. XL is only good because Pony fixed it. 2.x was a failure. SD3 was a failure, and that's all they have ever done.
>>101745626>Next question is will it be possible to NVLink 2 3090's together to train it?nope, and it's "illegal", Nvdia made it clear you aren't allowed to do that, or else they'll sue your ass, that's why you're forced to buy A100 :^)
>>101745642Good, give me an mi2xx then.
>>101745659>Nvdia made it clear you aren't allowedI hate NVIDIA I hate NVIDIA
>>10174562630 series is the last consumer card you can do that on, unless they decide to bring it back for 50xx but I heavily doubt it
>>101745666>I hate NVIDIAEveryone hate NVDIA anon, everyone...https://www.youtube.com/watch?v=UeU1WUb1q10
>>101745677I have 1 3090 and am strongly considering buyig a second if this proves feasible. Otherwise I'll be a GPU rentoid.
Comfy really seems to hate it when I change models. It crashes once before working when I load it again.
>>101745626>>101745677Also, according to simpletrainer dude, training on multiple cards works well for flux. So you probably don't even need to pool them.
>>101745666Why can't China make cards?
>>101745688>training on multiple cards works well for flux. So you probably don't even need to pool them.that would be slow though?
>>101745698I know nothing. He said it I am just parroting
>>101745679>t. kamala voter
>shills unironically trying to bill this garbage as "local Dall-3"
>>101745711it's better than dalle3 because it's uncensored >>101745631
>>101745687I have also found that if I kill ComfyUI, and run llama.cpp, then go back, ComfyUI would sometimes crash.
>mention a no no wordDOG
>>101745642That's why I want Intel to do it. By all rights Intel's software stack appears to be better put together for AI than AMD's, and Intel's backed into a corner right now and desperately needs a win. Offering a 48gb card at affordable prices, even if its compute is otherwise mediocre, would sell like fucking hotcakes and put them on the market for AI. And it's something where the hardware value is so good that people would MAKE it work, they'd find a way to make the damn thing run just to avoid going to ngreedia and dropping 8x as much per card.
5.65it/s, I did it boys
is 12 billion parameters a lot
>>101745738for you
>>101745734Great. I hope you die horribly.
>>101745738SD3 was 2B, and to be fair it's pretty sharp on some things
>>101745738>is 12 billion parameters a lot>SD1.5 is 0.75b>SDXL is 3.5b>SD3 is 8byeah it's a lot
>>101745749It's pretty clear how SD3 cant draw hands or a straight line for month things how important parameter size is.
So it's a lot. But is it enough?
>>101745747You can do it too, just drop your resolution way the fuck down. It works amazingly well at this size.
>>101745760name every concept you know, in alphabetical order
>>101745760>But is it enough?it is, look how good the outputs are
>>101745711huh
>>101745764>It works amazingly well at this size.It works at every size yeah, no more duplication shit you can see on SD models, flux really is an incredible model
>>101745779You think this DoFslop looks good?
>>101745687I've found comfy seems to just keep things loaded in VRAM even after you stop using them and you have to kill it to fix that. Maybe this has changed in updates idk, I had a months old version.
Any sequence of words that might work for removing the depth of field/blurriness? Or do the more realistic images just come out that way in flux.
>>101745791u cri erytiem
>>101745732>That's why I want Intel to do it.
>>101745801>Any sequence of words that might work for removing the depth of field/blurriness? Or do the more realistic images just come out that way in flux.use a negative prompt?https://reddit.com/r/StableDiffusion/comments/1el3tnq/want_to_use_negative_prompts_with_cfg_1_on_flux/https://reddit.com/r/StableDiffusion/comments/1ekgiw6/heres_a_hack_to_make_flux_better_at_prompt/
>>101745791I wonder if there is a way to control that.
>>101745811Have you tried "detailed," or "sharp" etc?
>>101745811>Drawback 3x sloweroofI'll learn to deal with the blur.
>>101745732You know that's what they should do. But you know what they're gonna do, right? They're gonna say FUCK the consumer and miss out on the opportunity to completely supplant Nvidia in the home GPU space and waste their time trying to out compete Nvidia in enterprise GPU sales. And when the whole AI bubble bursts due to everyone being oversold oversized GPUs, Nvida can just go back to its home GPU base and Intel will be left there dick in hand.
>>101745786Not every size. There's a breaking point
>>101745831WAIT WHAT????!~!!!!!WHAT IS THIS????!!!!!!~~~~~
>>101745831Fucking amazing actually
I love the shoes on the robot.
>flux makes something amazing>it's unreal
>>101745845I think this is the minimum viable resolution
>>101745764can you enlarge keeping that same image, tho?
>>101745831>>101745845YES!!!
>>101745859I will try, give me a few
>>101744381>06/08 2024No way this is random r-right?
>>101745869ok, here's a few
>>101745764>>101745859Yes
>>101745859>>101745869Interesting question.And does changing to shnell change the results?
>>101745878It's a real photo anon, this is actually a thread of real things, and your eyes (implanted by the Greys) show only fake things, except in this thread (BOOKMARK)
>>101745882you upscaled with SD, didn't you
>>101745878This is actual footage of Trump squatting in a convenience store from today.
>>101745786>>101745831>>101745856>All FLUX.1 model variants support a diverse range of aspect ratios and resolutions in 0.1 and 2.0 megapixels, as shown in the following example.https://blackforestlabs.ai/announcing-black-forest-labs/psa: any anons in this thread who have not read the official developer's announcement are encouraged to do so now, or at least skim it, it's not that long and it has pictures
>>101745711that 4chan's bedroom?
>>101745905you believe everything someone say? SAI said on their paper they were better than MJ based on their benchmark, do you seriously believe SD3 is better than MJ6 anon??
>>101745898Nope. Check it. files.catbox.moe zbl91z.png
>>101745738It should have always been the standard if not for vramlets
/ldg/ is going to hit the *post* limit due to (mostly on-topic) discussion
>>101745917NTA but I can use Flux and confirm it's very good.
>>101745884>does changing to shnell change the resultsYes.
>>101745946just to clear up any potential confusion: FLUX is a product of Black Forest Labs AI
>>101745962I know, I'm just more inclined to take BFL at their word because A their model works and B they have a history of actually making models that work.
>>101745923oh, basedlemme test
>>101745944fuck off with your saas slop
>>101745960>anonymoose hacking on a macit's so over, the feds are gonna get him
>>101745791dof fix is apparently just say "sharp photo of" at th front. At least that's what one youtuber says:https://youtu.be/1JtFK73K2sE?si=0od3SZsuTXmcUP5z&t=410
Neat Background: A lake with a boatForeground: A picnic tableSeems to work just fine without prose.
>>101745953Is it possible to get Shnell and fp8 to match? or no, they just diverge?
>>101745500I wish it showed her bra and panties but this is still impressive
The world needs this.
>>101746063I don't really understand your question. You can run Schnell at fp8.
>>101746077Should have a N on his hat :^)
>>101745923wait, this looks completely different to the regular flux workflow, where's all the flux nodes and all that? how do you set this one up?
Has anyone optimized flux yet?
>>101745944better start posting imgs then
I think I'm done looking at thighs trying to get this prompt to work.
>>101746085It's using comfys merge which is model fp8 and vae and weight all combined for simplicity.
>>101746082whoaaaaaa now I feel like my brain broke, hol up.I have...flux1-dev-fp8.safetensorsSo this means that I have dev, and it's been altered to be smaller?
>>101746063match? no. close? somewhat.>>101715965
>>101746098I like it.I think I have an idea how the no nudes thing came about. sleuthing.
>>101746123It looks like a slight change in guidance - but adjusting guidance doesn't get a match?
>>101746118first time using comfy, I'm only doing it because of flux, so I've no idea what that means
>>101746122Correct
>>101746029A woman standing with a smile on her face wearing shorts and boots full body shotBackground: A lake with a boatForeground: A picnic table5/5 gens with everything being where it should be besides the boots. Now if only the background wasn't blurry.
>fp8>fp16why does a fp12 doesn't exist? that would be the sweet spot between a model that will fill your whole 3090 and a model that is kinda small for your 3090
>>101746156try putting this at the front:a sharp photo of
I feel like certain words like "it" and "on" always get missed in text prompts
>>101746163Image gen virgins can't quant right. There hasn't been the need up until now.
>>101746171I feel like we need to combine tools, and be willing to use some photoshopping (gimp)
>>101745711Yeah it's not remotely as good stylistically but it's still fun
>>101746185I know how to use gimp nigger, I'd rather just keep genning. Sometimes gimp isn't a good solution either if you want well-done text in 3D space.
Are you all using Dev or Schnell?
>>101746219Dev
>>101746151All you need to know is if you don't have a 4090 use fp8 otherwise it will be slow as shit.
>>101746166Doesn't seem to help unfortunately. On the plus side I'm getting 16 second gens in Schnell on my 3060 so that's neat. The quality does seem lower though.
When ready... >>101746235>>101746235>>101746235
>>101746225that's not what I mean, I'm not a SDnoob, I mean the workflow is completely different from the flux workflowI can't even load the model because it doesn't see it, which is in the unet folder, and all the other flux nodes are missing toolike, I don't even, what is there a link to your workflow where it explains how to set it up?
>>101746232>16 second gensfp8 or fp16?and is that with 4 steps?
>>101746249https://comfyanonymous.github.io/ComfyUI_examples/flux/Make sure you grab https://huggingface.co/comfyanonymous/flux_text_encoders/tree/mainClip_I, fp16 and fp8 and put them in the right places. Grab either flux dev https://huggingface.co/black-forest-labs/FLUX.1-devor flux schnell https://huggingface.co/black-forest-labs/FLUX.1-schnell/blob/main/flux1-schnell.sftAnd put them in your ComfyUI\models\unet folder and NOT the checkpoint folder. The only thing that goes in the checkpoint folder is if you download his >Simple to use FP8 Checkpoint versionDownload his first two images from the first link and drag them into the comfyui window to load his workflow for examples. You might have to click the refesh button or the arrows to get it to properly find the models that you downloaded/put in the folders.
>>101746300I just have the fp8 checkpoint. It doesn't need the rest, if you just want to generate images.
>>101746265https://files.catbox.moe/cndtpu.pngYeah it's schnell with 4 steps. 100%| 4/4 [00:16<00:00, 4.07s/it]Requested to load AutoencodingEngineLoading 1 new modelPrompt executed in 19.16 secondsLooks like I was off by a few seconds.
>>101746300yes, I'm already using flux... I have everything installed, it works perfectlybut your workflow, is completely different, it looks like a SD comfy workflow, hende the confusion
>>101746334Oh my bad. I'm not him and didn't look up the chain all the way. I'll just leave that there in case someone lurking needs installation help
>>101746349no problemoI just want to know if people are upscaling with flux or not, every single image posted is in 1024, so I guess that a no
>>101746320I tried that and it gave me a bunch of errors...
>>101746362Did you get the comfyui manager? It can download missing nodes from other workflows. I'm also new to comfy but so far it's found everything that I was missing. Probably won't work if it's a personal node though.
>>101746384yup, I don't mean that there are nodes missing in his workflow, I mean the nodes of the flux workflow are not present in his workflow, which is mighty weirdis not that I don't have them installed, is that they're not part of the workflow he posted, which looks exactly like a regular SD workflow but somehow using the flux model
>>101745882>k>>101745764>You can do it too, just drop your resolution way the fuck down.That doesn't work for me. I get a SOMEWHAT different image if I change the size.>>101746371WHY?!!! I literally only added that file. How can I double-check?
>>101746406you added the flux model as a regular model and it worked?
>>101746406wait, comfy's fp8 model is like that toolet me try it again
>>101746402Forgot my meme arrow >I can't even load the model because it doesn't see it, which is in the unet folderAlright I loaded up his workflowLooks like he's using the safetensors version which doesn't go in the unet folder. It goes in the ComfyUI\models\checkpoints folder instead.This onehttps://huggingface.co/Comfy-Org/flux1-dev/blob/main/flux1-dev-fp8.safetensorsIf you already have that in there then I have no clue why it's not working. I couldn't run it because I'm missing the upscaler and don't really want to go find it.
>>101746455>>101746465exactlyif I try to run comfy's fp8 model workflow I get this, but I can run the fp8 model perfectly using the dev workflow...
>>101746413>>101746442I'm on ComfyUI. I literally just put the one file in the folder (and the other SD one, to get things working first).Then I drug the picture of the anime retard onto the middle of the ComfyUI tab.This is the full extent of my configuration of ComfyUI. I mean part of installing is installing pips but those aren't part of ComfyUI itself.
>>101746465>of the anime retardlmao, basedyeah, I think I'm doing the exact same thing, yet it's giving me an error for some reasonthe weird thing is that using the dev workflow everything works like a charm, both the fp16 and fp8 versions
https://civitai.com/images/22861951lmao
>>101746475did the sd model work for you? Let's make sure that there isn't something else broken.
>>101746544the model yes, but then it gives another error when it reaches the KSampler
>>101746544found thishttps://github.com/comfyanonymous/ComfyUI/issues/3693
it looks like comfy's flux fp8 workflow is a recycled SD3 workflow...
>>101746475
>>101746623this didn't work, btw >>101746565no one on jewtubew is using comfy's fp8 workflow either, but the dev workflow
>>101746623you asked the model to have a realistic background? that mix of style looks good!
>>101746640Basically "anime woman" and Background: Factory floor
>>101746565is the md5 okay?
>>101746681wait, hold the gayass furry fucking faggot fuck oncomfy's fp8 model is different from my fp8 model...my fp8 is 11Gb only, comfy's 17Gbthat must be the problem, his workflow uses a different fp8 model altogether than the fp8 model most people are using with the dev workflow
>>101746681>>101746688they're called exactly the same too, ffs
https://xcancel.com/jaguring1/status/1820254309558399416#mI wonder how he managed to get the first picture, that one has a different style that what we usualy have
>>101746711aha!!!!!!!!crazy, man!
>>101746750wonder if there's any difference in speed or qualityI'm downloading it now but my wifi is utter shit
>>101746718>different style Which part of it? How it's not as 'flat' as most anime pics from flux?
https://github.com/kijai/ComfyUI-CogVideoXWrapperOh shit did you see that, the CogVLM fags made a local text to video model, and it's not that bad
https://xcancel.com/doganuraldesign/status/1820120047379157365#mInteresting comparaison, as I excepted, MJ is better than flux, but let's not forget we're running a base model, we can make this shit even better
>>101746966>When making this comparison, I kept the scope broad and chose the best results after a few rerolls for both models.>However, given Midjourney's superior aesthetics and editing capabilities, FLUX won't surpass it soon.>If you think Midjourney is just about pretty pictures...>Then, MAYBE, now it has a stronger competitor.>But if you know they're a leading AI research lab, building hardware with 3D/Video support on the horizon...>Then, ABSOLUTELY NOT.This guy sounds like an absolute fanboy and his methodology is completely unscientific. He also seems completely unaware of the fact that he's comparing a base model with a relatively minor aesthetics bias to a finetuned model with a huge aesthetics bias, and using prompts tailored specifically for MJ with no attempt at taking advantage of Flux's dual text encoders.
>>101746966What an absolute retard kek
I have discovered that Flux uses some kind of advanced trick to try to scrub out p*rn.working on low resolutions, play with guidance. I put cfg on 3, 5-9. This is at around 400x400 or less.It's like if it can think enough it goes through an unthink cycle.I believe there are multiple factors at work. For one thing, I think some images it was trained on were blurred out.
>>101747105>a base model with a relatively minor aesthetics biasyou're joking? when you ask for an anime picture it always gives the same style
New bread if you missed it >>101746235>>101746235>>101746235
>>101747105>with no attempt at taking advantage of Flux's dual text encoders.how exactly do you take advantage of those?
>>101747139If you don't specify a style it should give you a generic style, that is intended behavior. What it doesn't do is beautify it without you asking it to, or at least not as much as MJ
>>101747154>If you don't specify a style it should give you a generic style, that is intended behavior.no, if it has no bias it should give a variety of styles, not just one, that alone is the sign it has been finetuned a bit
>>101747153An obvious place to start would be giving old school tag based version of the prompt to CLIP and natural language version of the prompt to T5. But more investigation needs to be done
My current game is jailbreaking Flux.
