Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101770020>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
>mfw
it kinda knows what a straitjacket is, but not really
comfy is shit. flux on auto1111 when?
>>101774535two more weeks
>>101774535comfy is best.auto1111 hopefully NEVER EVER
>>101774535never, except unironically. only via some third party extension and no official support is my bet. just use swarmui if you are getting filtered by connecting color coded strings together
>>101774485twonaters
>>101774535When the dev stops his 24/7 gaming session.
>>101774584swarm is far from fleshed out enough to be worth is. aspect ratio values don't even function properly in the sidebar
>>101774497Yeah at this point it seems like someone who just does his stuff silently without entering useless dramas is a rarity;
Can someone share an example workflow with high cfg and the cosine thing? I have no idea what it is.
>>101774584Auto1111 is painful as fuck to develop with.
>>101774535I agree. Workflows are for promptlets.You require 6 hours of workflow setup.I require 6 minutes of prompting and lora assembly.We are not the same.
>>101774636I just get a good workflow, pretty sure I can get one that does what a1111 does exactly.
>>101774288Why does Flux need two generals?
passed midnight
>>101774535relax catjak>>101774382every block .5 except the last, final block 1.0 (all Flux)
>>101774661>tile walls for an insane personDamn, they want her to kill herself
>>101774652skill issue
>>101773896don't go to deep into the forest
>>101774535Swarm exists, why the fuck would you need auto at this point?
If I want to compare two workflows, outside of launching two sessions, is there any only website I can upload a workflow to so I can have both my local and it up at the same time?
>>101774699why the the hell does this shit feel so much more detailed despite the same 1024 res as sd3? and i cant even say sd3 gens on their own felt particularly lacking in detail to me when i was genning.
where in workflow on ComfyUI can i put image preview? Only place i see with that light blue connector for image is after my VAE decode at the end
>>101774739awewsome
>>101774535>a1111 overhead + FluxLMAO
back at this seems to work better talking about the girl first
>>101774739I like how flux anime looks like it came out of a TV anime or a promotional piece straight from Japan. Not the usual Chinese sloppafied digital art style we usually get with stable diffusion.
>>101774848I think like all thing stable diffusion, the lower down in the prompt something is, the more likely it is to just be forgotten.
>>10177485290% of the people that wrote the SD3 paper left the company to make Flux, this IS Stable Diffusion
>>101774852it seems like it was trained on a lot of anime. if i say 80s style it looks more oldschool. or maybe i'm just imaginging things
>>101774750yes it's very sharp and clean, even the entrails and offal looks like plastic lel maybe a side effect of making the text so clear
6 times as many parameters. they almost entirely busted their load releasing this model.SD3 is 2 billion parameters, Flux is 12 billion.
>>101774862I know, but the anime models are full on slopped with over detailed fan art. I'm sure flux will end up that way too.>>101774872It has a few variations but they're all pretty similar desu
>>101774861So the opposite of LLMs then?
>>101774916By lower down I mean closer to the bottom. Do LLMs prefer things down the bottom?
>>101774925For LLMs, the lower the text the higher in priority/memory it is.So yeah opposite.
>>101774904That's still nothing compared to LLMs' parameter count. The only issue is really the computation load per parameter is much higher and quantization in image generation is barely being done or done at all. That may have to change but I don't see anything other than 4 bit quantization being done after FP8, there's nothing like GGUF or EXL2 and it's not clear you can do that kind of quantization here.
>>101774872try "by Leiji Matsumoto" for some real 80s anime
>>101774935And no support for multigpu either for some reason.
Why official Wishlist for flux>Finetune so it knows what bob and vageen is and will show it on request>LoRA support, preferably trainable on a single 24gb (will buy a second 24gb card if training on that proves feasible)>Controlnets, particularly depth, openpose, IPadaptor and canny. Anything else is a bonus.
>>101774952>Finetune so it knows what bob and vageen is and will show it on request+ styles, artists, characters, people
Bigma status?
>>101774968Short of a massive pony sized finetune I don't see that happening. And the pony guy is a massive faggot so I don't think he'll touch flux. Probably going to have to make do with LoRAs to get the styles you want.
>>1017749522 and 3 are already met. imagine if you were in the discord t.co/1AkeueFGc5 t.co/RYOmJ9vXETthere are already 3 loras including an IKEA instruction style lora and a realism lora
>>101775000I know but I can dream.
can anyone give me a catbox sample of how to do cfg with floox? :)
>>101774952>Controlnetswas released today>https://huggingface.co/XLabs-AI/flux-controlnet-canny/tree/main>LoRA support, preferably trainable on a single 24gbis said to work https://twitter.com/ostrisai/status/1820829417595076623
What kind of resources would it hypothetically take to finetune a flux model? Do we even know?(reposting from previous thread, because I want to know and am a fag)
>>101775035yeah. this one has CFG and negshttps://files.catbox.moe/m9mv31.pngignore the model merging, just delete all the extra loaders and only use the diffusers loader>>101775036slow
Is there any reason to use flux dev over schnell?
>>101775035here you go:https://files.catbox.moe/p33h6q.png
if you only knew how good things could really be
>>101775064Dev looks better and does text better. Schnell just looks really good for what it is.
>>101775062>>101775066thanks frens
>>101775064i know who you are>>101775104yw
>>101774904can you make a warthog with 4 eyes? Hasn't worked on any engine ive tried.
>leave general for just a month>there's another prompting revolution apparentlyI love this
>>101775064>>101775092>>101775117meant to reply here oops>>101775119will try
>>101775122imggen has risen again prompters UNITE
this one's to poor to buy a real videocard
>>101774936gonna try thanks (this is an old one that did not contain Leiji in the prompt)
give me your fucking vram
>>101774422pony
>>101775119not even a warthog really. adjusting a bit and trying again
>>101775198Looks more like a bush pig. weird to see one on a beach though.
>>101774936came out good, not quite what i wanted but i like the style
>>101775119probably difficult to prompt, can make nice spider warthogs tho
>>101775230Every time I get my hands on a new engine I try it. But I havent found one that can do it yet. It's a poster I had in my room when I was a kid. It stuck with me somehow
>>101775258Is there an engine that can do deformity, yet? Like shriveled hands or arms?
>>101775249>all prompt and no vram makes jack a dull boy>all prompt and no vram makes jack a dull boy>all prompt and no vram makes jack a dull boy>all prompt and no vram makes jack a dull boy
>>101775304very nice and detailed
>>101775268>>101775317catbox? maybe I will try this model
>>101775337Flux doesn't know what flat Earth is.>it's ogre
>>101775367its joever
been using howzoonga to create some stuff recently, its an ai gf app but it does the job lol
>>101775349Flux, it's a must. The other models can still do sweet amazing things, check this thing out.
The day is coming bros. The day I can describe every detail of my futa wife's massive veiny 3 foot weapon of a cock and it will generate exactly what I asked for.
>>101775408I mean, it was always just a matter of time anyway
>>101775415hey, how'd you do that? I tried to put flat earth in space, and it kept making globes.
1080ti for lief
>>101775349sorry I was messing around with prompts birds eye view and moebious and gurney where in there I gen using online versions
>>101775417flat earth on a arabesque golden table carved with jewels, memento mori, dark, smoke
>>101775408in full interactive VR with touch.
>>101775379flux always makes hot ladies
>>101774944To be fair, it wasn't needed really until now. It should be completely possible but you have to make the diffusion deterministic so every GPU can get a portion of the offloaded generation and produce the same results.
>>101775367>>101775370don't you worry all is well
>>101775480Yeah, it has baked in tarting
>>101775503bet the catfishers are losing their mind over the potential
>>101775498>don't you worry all is wellGreat, what's your prompt, I clearly did something wrong, not sure what's up.
>>101775519I did>biblical depiction of a flat earth, disc with waterfalls on the side surrounded by empty space, you can see continents, oceans on the flat discprobably can be optimized, maybe with negatives like .. just added "globe" to negatives for this one (not the previous one)
>>101775519this is even better:>depiction of a flat earth, disc with waterfalls on the side surrounded by empty space, you can see continents and oceans on the flat disc, the edge of the disc has icy regions, the background is just empty space and stars
>>101775590looks like a 1080ti to me so i approve
>>101775600
>>101775549>>>I think Flux is trolling me>flat earth disk spinning in space
>>101774288>https://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellwhat's the limit on this?
>>101775645i like it, good troll
>>101775600It's the 101090. Says so on the board. Here's the upgraded typo verison
>>101775645you have to give it more information, only mention earth to get the continents right, then describe the disc as >>101775594 also it helps to forbid "globe" with the negatives hack as in this workflow >>101775066 add globe to negatives, use the prompt above and you should get what I did
>>101775655there isnt, the more you use the lower your priority becomes, until eventually it takes 24h+ per image
>>101775643Good, but the water is leaking out of the ice wall.
>>101775661okay, pretty nice, thanks
Anyone done any tests on this shit yet?https://civitai.com/search/models?baseModel=Flux.1%20D&baseModel=Flux.1%20S&modelType=Checkpoint&modelType=DoRA&modelType=LORA&sortBy=models_v9They came so quickly I can't image it's anything.
>>101775687Your GPU spent 40+ seconds burning, whirring and generating enough heat to fry and egg for you to produce this shit.
>>101775697doesn't even make a sound
>>101775693I'd wait for things to settle down a bit. The day Flux came out retards were trying to merge it with pony and all kinds of dumb shut. It's basically just a catalog of failed projects right now.
>>101775655limit to what? schnell just does 4 passes where dev does 20dev works on my 1080ti too
>>101775697You make shit eggs
>>101774288>Flux model is 22+ GBChrist Almighty...... I'm wondering if the machine I use can even run that... How much GPU VRAM is necessary to run inference with this in a timely manner?
Siding with this guy for now:>>101775697>>10177571255c edge?
>>101775666kek, I think of Discworld when I think of flat earth, there you have waterfalls, also its carried on the back of a turtle that flies thru space..but I guess this the flat earth society official version:>depiction of a flat earth, disc surrounded by empty space, you can see continents and oceans on the flat disc, the edge of the disc has icy regions, the background is just empty space and stars, empty outer space background
>>10177571512GB VRAM with 32GB RAM is the safe minimum.
>>101774535I really don't understand why he hasn't implemented it. I've been running it without comfy since release
>>101775732How long does it typically take to generate an image compared to SDXL or SD 1.5?
>>101775732I have that exact setupfor some reason it insists on going to low VRAM mode, regardless of if I try fp8. seems my gpu can't do fp8 calculations (using a 3060)
>>101775715I am using the ComfyUI fp8 one, like seemingly everyone has. Link in op. NOT the huggingface one, the other one, the comfyanonymous one.fp8 is in the middle. I can do it on a 6950xt, which is not supposed to be happening. But, apparently a 1080ti is just as good lmao (well, at THIS)
>>101775745depends on your hardware, on a 4090 a 1024x1024 on FLUX.dev is 15-20 seconds, FLUX.schnell less than 10 seconds
>>101775738>without comfy since releaseon what?
>>101775732>32GB RAM is the safe minimumPeople are unaware of how critical regular ram is. 16gb, and it is so slow, I don't know how slow, it basically doesn't work.
>>101775766Difference between .dev and .schnell?
>>101775790top kek
>>101775790someone said put it on a gradient background.idk why people are trying to sex up being a Republican. True, Republicans don't really want to be called weird, but the #1 enemy of the RNC is dullness. Conservatism itself is basically preservation of something.
>>101775745For me, 4080 16GB VRAM 32GB RAM at 50 steps FP8 dev takes 60 seconds per 1024x1024 image (including pre-process and decode), around 1.2 s/itSDXL is 4.5it/s and 25 seconds per image.1.5 is roughly the same at that resolution but functionally longer as I need to do an upscale step
>>101775779dev needs more iterations (20+) and is more precise, schnell is like a hyper sdxl model and converges in 4-6 steps, but the output is kinda limited .. schnell is the free and open source one, dev is the free but non-commercial model (there is also the master model FLUX.pro, both schnell and dev are distilled versions of pro)
>>101775754so is this a problem?
>>101775732AMDchads we're in
>>101775804Have people figured out how to train LoRAs for flux yet? How much RAM would you typically need for that?>>101775814The fact you mentioned they are free and open source implies there are some restrictions in some ways. Are they doing that shit stability try to do with SD3 where they try to claim they own all derivative works made by the model? Or are you referring to the fact that the training info is open source?
>>101775816NTA but definitely. Using your CPU is slow as fuck.
>>101775832Sorry I should have specified NVidia only but I thought it was a safe assumption.
>>101775841>Have people figured out how to train LoRAs for flux yet?yes>How much RAM would you typically need for that?the more the better
>>101775858Cat box for this please. I want to try something
>>101775730>The Earth is a disk, akin to a map, slightly curved, over which has a clear shell which is the firmament keeping the atmosphere in. All along the sides, there is a icy mountain range, which contains the seas.Mine doesn't work. It's a ball, I canceled on preview.
Guys, I'm about to pull the trigger on second 3090. Talk me out of it.
>>101775868>I want to try somethinglike what?
>>101775841>The fact you mentioned they are free and open source implies there are some restrictions in some ways. Are they doing that shit stability try to do with SD3 where they try to claim they own all derivative works made by the model? Or are you referring to the fact that the training info is open source?FLUX.schnell is apache 2.0 .. more free is nearly not possibleFLUX.dev is free and open source but non-commercial use only, even if you fine tune it, thats why some of the coomer fine tuners are on the fenceFLUX.pro is completly closed source and only available on per API directly from Black Forest labs
>>101775880go for it, buy it.you need it, you want it. give in to your prompt addiction.
>>101775851you think its only using CPU because it says LOWVRAM mode?it sure uses the fuck outta my GPU on LOWVRAM mode
>>101775851for the flux model yeah, for the t5 model not really
>>101775697based
>>101775900it's just not being utilized for meI checked usage and stuff, but nothing
Anyone having ram issues with flux on amd using comfyui? I've tried both dev and schnell and also both fp8 versions. Every time my ram peaks out and i get a crash. I can run other models fine on comfyui. I've also used different variables for fp8 --lowvram --normalvram and consistently get the same issue every time.
>>101775697
>>101775594tried your exact prompt.>resolution matters(!!!)
>>101775754Do you need a 4000 gen gpu to use the model in float8?
>>101775885>more free is nearly not possibleWhy not? >>101775885>but non-commercial use only, even if you fine tune it, thats why some of the coomer fine tuners are on the fenceDoesn't that mean you aren't allowed to make money from outputs? Why would local-gen compete care about this? Wouldn't this also conflict with an online AI art generation site's business model since they technically make money from the use of any and all models there (assuming it's available for generation, like all three versions are on Civitai as of writing this)>FLUX.pro is completly closed source and only available on per API directly from Black Forest labsDo they make a lot of money from doing this?
>>101775983It helps, but no. YMMV especially with a XX60 card.You should probably be only trying Schnell on that. Don't worry though, Schnell is miles ahead of SD3 so you can still produce good shit.
>>101775754my 6950 is stuck in low vram mode, because of how rocm works (I think, ie, they say it uses some gpu ram to make it work). Still generating.>>101775982>>101775594Your exact prompt, with the correct resolution.
>>101776005>Schnell is miles ahead of SD3This is true, however sd can do amazing things.
>>101775955i'm going to ask a dumb question because mine works fine and uses gpu on lowvram modedid you start comfyui with the cpu bat instead of gpu?
Keep in mind guys, you can gen at 512x512, even as low as 128x128 and still produce very useable results.
>>101776037I'm running this in endeavour os (basically arch) so I start it with: HSA_OVERRIDE_GFX_VERSION=10.3.0 python main.py. I have ran different SD models while checking my system monitor and it runs fine and pulls my GPU resources. I'm running and 6700xt.
>>101776001>Why not?Apache 2.0 license is one of the most open license models>>101776001>Doesn't that mean you aren't allowed to make money from outputsprobably? I am not a lawyer .. but at the very least you can not use it and sell your finetuned models or make them paywalled, or put on a webservice where ppl pay for its usage>>101776001>Do they make a lot of money from doing this?no idea, it was released not even a week ago, so no info about it is out>>101776013hoooray! success! seems resolution really matters
>>101776013>>101775594BET YOU DIDN'T EXPECT THIS!>same prompt>swapped 768 and 1280 in dimensions
>>101776029>sd 256x256vs >>101776079>flux 256x256
>>101776094lol.. flat and globe earth!
>>101776079
>>101774646imagine, you get a 4090 and a jappy school girl UOOOOOOOOOOOH
>>101776092>put on a webservice where ppl pay for its usageIsn't civitai literally doing that right now?
>>101776119
>>101776114
>>101776005Alright, thank you!
>>101776129What prompt do you use for that peephole effect?
>>10177569740 seconds to gen, no whirring, gpu temp hanging out at 46 degrees, spiked up to about 66 degrees for 10 seconds or so... them some shitty eggs.. i wouldn't eat them
>>101776145>House doorbell security cam footage photo. Fisheye lens effect creates rounded distortion
check out my ninja skills, gosh!
>my old domain names are estimated at ~20k-30k according to different websites>can't be fucked to sell them to get a few 4090shello, I'm a retard
>>101776171>at ~20k-30kwhy not get some A100s for that cash?
>>101775767using my own scripts for it
>>101776129no I prefer the 4090 and a jappy little girl
>>101776186How does one even buy an A100?
>>101776089like i said stupid question but just wanted to make sure it wasn't a silly mistake causing your issues
>>101776171Are those domain names at least useful to you if you're hoarding them?
>>101776199you order one?>https://www.newegg.com/p/pl?d=A100+Nvidia
>>101776210i'm just waiting for them to rocket to the moon like bitcoin
>>101776199you order them online like anything else?
mfw i have a 4090
>>101776219
>>101776186I haven't thought about it very much. that's a possibility too, I guess. I really don't know if that's the "true" price of them>>101776210well, they've gone up in price significantly as the terms related to them got more popular. my laziness has actually paid off>>101776216 was not me
>>101776237
>>101776269
>>101776237let me see those sweet render timesso me and my 1080 can cry
>>101776241
>>101776306
>>101776249you should go for it bro.
>>101776323jesus thats 10+ times faster than me and you're doing 30 so way more
>>101776371i gave a friend my 1080 ti like 3 years ago when I got a 2070
DynamicThresholdingFullWhat's it for, in the Negative example?Also, why is my time like 7x slower for Negative?
>>101776323
>>101776420kek'd
>>101776358maybe I shouldthanks, anon
>>101776323Why 30 iterations???
>>101776420Very nice.
When will the flux honeymoon phase end?
>>101776434why not
>>101776411Someone gave my poor ass this one when they got a 3080
>>101776450when I get bored
>>101776452idk, does it help?Also, what speed do you get with:>>101775066(I changed it so I can use fp8)
>>101776407
>>101776457When the flux finetunes come out. I don't think flux is going anywhere. The community are bending over backwards to give it tooling.
>>101776465>the only person who can honestly say powerpoint saved his life
>>101776430based
>>101776468i dunno.. it depends on the sampler i think.. the dpmfast pics are shit unless i have more steps
>>101776520>he bought a slutbot instead of a maidbotWhat a fool, just look at him!
>>101776468took 1 min 11 seconds
>>101776568fried subtitles
Some of the Flux images you guys have been making are almost decent looking but why the hell do I keep seeing people saying it's local Midjourney? It's clearly not as good looking as Midjourney.
i think i can kinda make out the word vram maybe
>>101776610ask them...?
>>101776619I am!!!
>>101775119>can you make a warthog with 4 eyes?
>>101776558Very nice. For whatever reason, I have to switch to Euler, simple, to get it to take twice as long, on my 6950xt.
>>101776642(I get 14.88 sits, and so <30 sits with Negative)>>101776638daw
Can you do weights like (fag:1.3) in the text you send to the clip_l encoder?
>>101776610It's definitely not local midjourney it's actually something entirely new. It's an image generator that gives you what you describe and looks decent doing so.
so hit and miss
>>101776610MJ is dogshit, it's only "good" because the gacha sometimes creates good looking images ignoring like 2/3 of your prompt.
>>101776729cute
bake?
>>101776783not mine, someone did this yesterday and posted catbox
where can I make infinite memes for free with flux? i have a 1080ti from when I wasn't poor and it takes 3 minutes to make a 512 by 512 with schnell and anything larger crashes. I don't care if it takes long still I just wanna make shit but all these cringe sites cut you off real quick
>>101776833just buy a 3090 bro
>>101776833https://huggingface.co/spaces/black-forest-labs/FLUX.1-dev doesn't cut you off
>>101776833rob a computer supply store like everyone else.
>>101776840yes it does i am on like 30m wait time right now there :'( i made like 8 images before hitting that
>>101776864you can also post your prompt ideas here and maybe some of the kind Anons with 3090s and 4090s and A100s will do it for you.
>>101776840has flux pro leaked?
>>101776864if you're not worried about wait times and you're on a limited budget your best bet might be to just buy a bunch of ram and stick it in your pc. at least then your shit won't crash (probably)
>>101776909no, why would it? It's only on BFL's own servers, Replicate etc use it through BFL's own api, you can easily see that because for schnell/dev Replicate returns image URLs that are hosted by them, but for Pro you get bfl's own domain from Azure storage or some shit.
>>101776909could anyone even run this shit without an A100?
>>101775900Aren't you the anon with the 3GB 1080 Ti?
>>101776953Good question, if you read the announcement they actually don't say how big the Pro version is, they just say that both Dev and Schnell are 12B models distilled from Pro.
>>101776953I have no idea, but I suspect it's a good deal bigger than what would fit in a 24gb GPU at fp16, but would also fit rather comfortably in a 24gb GPU at fp8
>>101776925https://en.wikipedia.org/wiki/Jim_BellI'm sure anons could arrange a bribe using Ether, it's gotta be possible (like a voting-based determination of where donated funds later go, as long as the leaker encrypts the file with his signature, then you'll know the associated wallet is his)
I checked dev vs pro on the API, and pro isn't that much better, and in fact e.g. renders text worse than Pro. is it really that better?
>>101777029>>101777029>>101777029