Death to Vramlets Discussion of free and open source text-to-image modelsPrevious /ldg/ bred : >>101744342>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
I made the collage twice. What do I win?
>>101746270Cancer
>>101746272N-no.
>>101746287Don't worry, there's a cure
>>101746249Catbox an image from your workflow. I will set it up for upscaling for you and catbox it back. >>101746235kek at my tiny miku in the middle
>>101746270>What do I win?Our gratitude :3
Flux Pro is absolutely NOT better than DALL-E 3 - neutered and censored as it currently is at Image Create - and it can't hold a candle to the first release from last year before all safety faggotry. Is there any research or evals on prompt upscaling vs "tag prompting"?
>>101746342nice reference anon https://youtu.be/3sVBK7K2X6I?t=50
>>101746342lol
>>101746342that's insane how good it is at text, like it just get it, what a leap forward everything we've ever known at this point
Still gotta fill up the first thread. Don't get this one nuked.
Why does Flux have that Stable Diffusion vibe? Like it's aura is totally stability pilled. Isn't it really new and stuff? It's giving me the ick.
https://github.com/comfyanonymous/ComfyUI_TensorRT/commit/66f5bf191c8844904adcab436d2ab7c236046aa7So has anyone tried building the flux wheel or what?
>>101746659how does that work? it'll make the gen faster?
>>101746235So premature you fucking retard
>>101746680Much faster, quieter, more efficient for RTX owners.
>>101746724oh nice! so... how do I make it work?
>>101746731Install the extension and follow the instructions on the Github to build the wheel, since Flux is big it may take several hours.
>>101746755can it build from the fp8 dev instead?
>>101746768no
>>101746768No idea, read >https://github.com/comfyanonymous/ComfyUI_TensorRT/commit/76cdd93680c536a6f5683f60539eeabf27b39775
>>101746774:(
>>101746778>Flux needs to be bf16.but we only have fp16 models?
>>101746774once you've made a TensorRT file, you can't load it in fp8_e4 like on a regular file right?
>>101746755Can someone just build it and share the file on torrent?
>>101746788Flux is bf16
>>101746862everyone said fp16 since flux came out though kek
I love how retarded and overkill the cooler on the 4080s
>>101746841Yes, it's different for every GPU and CUDA configuration though
>>101746296is this bigma
Early bake really took the wind out of it
>>101746968nah the amerimutts are going to sleep, happened yesterday too
have we got something new that's not flux yet? not a single flux gen is aesthetically appealing
>>101746996anon thought >>101741569 op was cool
>>101746996Ai generation isn't aesthetically appealingPick up a pencil
official pixart bigma and hunyuan finetune waiting room
>>101746996pixart team said they'll try to release bigma on september, besides that nothing else seems to be on the horizon.
is there a way to save the workflow on cumfyUI that doesn't force me to type the fucking name each time like a literal fucking nigger?
>>101747073what do you need to save it for so often? all of your gens have the workflows embedded in them automatically
>>101747084because it arouses me sexuallyis there or not?
any good gens lately, anon?
>>101747099edge on windows assigns unique names to each workflow what more could you want you're welcome
>>101747148you could've just said "no"I don't mind socializing, tho
>>101746996quantify "aesthetically appealing" because that is subjective as fuck>>101747029picrel also kys
>>101747113nsfw at present.I am attempting to defeat the protections they put in place against the booby.
If I want to try flux, comfy is the only way? I have an old version of forge installed and an even older a1111 from the 1.5 days.
I have discovered there's a face-enhance layer somehow.what is the catbox rule on blurred porn? I didn't blur it, Flux did.So like at the first levels of the thing, it has a blurry knowledge of porn. But then there is the face write layer. And that happens, and would anyway, but then there is I think more enhance, but then it does clothes write, I think. I don't know how you knock face write and clothes write out of commission
>>101747051back into my coma for a month, see you all then
>>101745923Afaik the gen time will be the same with straight to 1024x method. You will only gain time if you upscale it with lighter models like SD 1.5 or XL.
>>101747205You can post a link on catbox to a woman spreading her asshole if you want to.
>>101747189i somehow got flux to show a little nipple or areola, previously i was under the impression that it thought females had nothing but skin over the whole breastdon't want to risk someone saying its nsfw though or i'd post
>>101747200>If I want to try flux, comfy is the only way?yes, unfortunately
:^) It's a game.I think it's like(lowres main gen)LOOP:-> enlarge, blur-> enhance X-> upscale detail addX={face, hands, clothes}idk, something like that.
>>101747200Trying locally? Yes, currently just comfy>A11112 weeksTrying online? Check OP. Several options.
>>101747215https://files.catbox.moe/clnq7v.png>>101747218They seem to always build on top of other already trained models with this stuff. Could be wrong.
Apple question: is the M4 coming out, and will it have ML enhancement?
kek
>>101747237>https://files.catbox.moe/clnq7v.pngYour guidance is too low. Happens to me too on non-porn.
>>101747221>>101747232Thanks, comfy it is then.
AHHHHHI meant to say STEPS. not cfg, if I was unclear.At step 9, it appears clothing is enhanced. At least maybe. It may depend on other factors.
>>101747251what did you say anon? lol
>>101747254Look again bro. The face is painted on top. that's what I want you to see.
>>101747262>At step 9, it appears clothing is enhancedof course, it has been censured, they just make the model believe that "naked" means "a human with underwear" or some shit
>>101747264someone made a reddit-tier joke in response to a serious post and I told them to kill themselves
>>101747237it really likes to force clothes on and even under clothes when you say see through or mesh or whatevertry this"wearing nothing facing towards the camera squeezing her breasts" maybe i got lucky though
>>101747272Oh I see what you mean.
>>101747277kek
>>101746996Is this aesthetically appealing?
>>101747276step 6 has tits still, though this may somehow depend.step 7 is titty erase mode.We want to bypass steps.
>>101747297>We want to bypass steps.that's not how it work, it will do that process anyway, whether it's at steps 7 or an another number
>>101747262>I meant to say STEPS. not cfgbut you didn't say cfg? You said layers.>>101747297that's not how things work
>>101747293Yes. However, I suspect anon is trolling
>>101746996Is this?
>>101747306It's absolutely how Flux works :^)
>>101747335No, Flux likes to really pump up the hos, and I am investigating nudegate rn, but I will investigate milfgate next.
>>101747321I know, but I just want an excuse to post some of the gens I really liked, and taking the bait is fun.
>The SunAccording to FLUX
>>101747335>a womanGay
>>101747254Also, I strongly suspect this is an error in vae.
>>101747365Flux hasn't seen a single nipple in its short life.
>>101747372I It's seen more than SD3.
I'M WRONG!!!!IT ADDS ALL FIXERS AT ONCEIT IS A FIXER STEPBWAHAHAHAHAHAHAHAHAHAHAH
>>101747394Who asked?
>>101747403anonie, you don't know enough about how these things workthere are no fixer steps, that's not how any of this works
>>101747413Focus on fixing your model Lykon it's pathetic at this point.
>>101747413
>>101747447is she preggers?
>>101747452I didn't specify anything of that sort, the model made her pregnant by itself
>>101747474nips
>>101747372Flux has a titty eraser step. or a titty eraser substep???https://files.catbox.moe/isntdy.png^ Flux knowsThat's step 6.But it erases them in step 7.
>>101747495those are just the same shitty nipples you can get without skipping any stepsI'm sorry, anon, it just doesn't know what nipples look like.
>RTX 5070 is still gonna be 12 GBNIGGER
>>101747452>>101747461And now I am going to make her pregnant intentionally>>101747495I have genned shit loads of nips on Flux what are you even on about.
>>101747506Could be it has only a vague idea.But there IS a nipple eraser:>>101747508compare to:>>101747495
>>101747495>>101747474>>101747508Anon...https://files.catbox.moe/w70kv0.png
>>101747509desu flux is working fine on my 1080ti just takes a bit >Prompt executed in 369.29 seconds
>>101747529whats sort of prompt is actually making them topless?i always get some sort of top
>>101747529Great job!Now we need to disable the whoremaker. and muscletoner.At least it shouldn't be on all the time.
>>101747532shnell?
FOR THOSE SAYING THERE ARE NO SPECIAL STEPSlook again at this:https://files.catbox.moe/clnq7v.pngClearly the face has received enhancement.
>>101747553YES!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!I love Flux, but what Flux is, it's a system of loras.
any nudity and female celebrity likeness comes from what CLIP knows, which is little
>>101747562no devthere really isn't anything in this world the 1080ti can't do except ray tracingi mean obviously it has to be at 1080p or below but thats why its called the 1080tishow me a game that I can't run
>>101747529Anon is trolling, anon. Report and ignore
We are so back.
>>101747592so is the guy saying a 1080 is that fast imo. I have seen other strange claims that are impossible.
uhh i guess this is what it thinks a nip ishttps://files.catbox.moe/ywb3gk.png
>>101747545Prompt is literally just "photo of a naked"Seed: 264852884050485I queued up 10 and hit go.I was exploring latent space>>101747574You specified an image of a person so it re-arranged the pixels into a persons face. What exactly do you think this is showing us?
>>101747597Very clean text
>>101747632Correct. It generates bee stings instead of nips. You have just discovered what the rest of us discovered 3 days ago. Here is my 2nd ever gen on Fluxhttps://files.catbox.moe/g3o9jq.png
Here is my plan. Learn to delete the loras of Flux, and then generate images in 1/3rd the time :^)side benefit: ho-ification removed
>>101747632did you try seeing which step adds that?
>>101747619>369.29 seconds>that fastanon what
>>101747682one legged man using a wolf as support
>>101747619>>101747669i never said it was fast i just said it works
>>101747669same as my 6950xt.
>>101747725oops meant to screenshot it
>>101747667https://files.catbox.moe/qsexdz.jpgAnon you are just wrong. Whatever you think is happening, it's not that.
>>101747736Sure, I got it a little wrong.It's doing the enhance face sequence for each step. It has enhance face and basically tone up. It's like a whole set of boosters, that have to get triggered.
>>101747755no
>>101747736You do know it is constantly overwriting itself, right?
>>101747755>>101747765I am very interested to learn how you think latent diffusion works.
anyone's tested ultimate sd upscale with flux?
>>101747730Isn't 6950xt much better than a 1080ti specs wise? Is it because no CUDA?
>>101747765>>101747773FLUX has noise guidance right? So could the trained guidance be weighted towards certain negatives and making it lean toward censoring?I have definitely noticed the attention moving around the image during generation (shows up like a hotspot)
>>101747831It isn't censoring, Anon. There's no censoring step. This is just what it thinks nipples look like, because it was trained to think nipples look like that.And I use think in the very loosest sense of the word.
>>101747773Why do you assume they are doing exactly what you think they are doing?>>101747844Both can be true. It can both not know what highres tits look like AND be self-censoring.Again, you know that the image is being edited, and if you adjust the step you can see a moment in time as it's being edited. You know this, right?
Anons, tf is this?
>>101747844Also, they have some fun stuff in here.I found the first (?) Flux easter egg: anus is trained on a bug lmao
>>101747854This is how latent diffusion works. https://www.tiktok.com/@pebblepicassos/video/7249998510269975813
>>101747870tiktokfuck off, really
>>101747844I get that there is missing data but the noise still seems to be pulled in a particular direction in a way I don't recall seeing in other censored models
No wonder America is falling behind China, we got a retard here who is Tiktok educated.
Is Flux an auto-Lora?
>>101747335Tits should be much bigger
>>101747895It was the first link on google that illustrated my point. Where would you prefer I link from?Here's leddithttps://www.reddit.com/r/oddlysatisfying/comments/14plfld/portrait_of_a_cat_out_of_rocks/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_buttonPoint is, you feed noise, you say "This is a dog" and it moves around the pixels to what it understands a dog looks like. Then it does that x many times for however many steps you chose.
>>101747938y-y-you d-disagreed with t-teacher
>>101747898Is there a way to dump a particular moment, and manually make changes, then continue, and do that again etc?
>>101747988you could just diffuse to certain step, save the img, then change said image do changes whatever/edit the prompt and finally in a different workflow use that image as input vae encode (apply noise or not) and run additional steps
>>101747988Yes you can stop generating at a particular step and do something else with the latent. KSampler (Advanced) in Comfy will let you do this.
>>101747938step 5 when requesting 10 steps does not equal step 5 when requesting 20 steps. the magnitude of the step delta is pre-determined as a function of total_steps. you can't take a 10-step gen and "choose" 10 more steps
>>101747988SaveLatent and LoadLatent nodes (in comfy)
>>101748048What the fuck are you talking about?You request 50 steps. You feed the latent into a ksampler, start at step 0 and generate until step 20, you take the latent out, do something else, then feed it back into another ksampler where you start at step 20 and end at step 50.
>>101748074ya thats the smart way to do it
>>101748048Shit I thought you were replying to a different comment. Now I am even more confused.
>>101747976
how long til we can expect to have a cpp version of flux that can run on CPU ?
>>101748164>diffusion on the CPUare you a masochist, anon?
How exactly does one boomer proompt efficiently? Do you write a long and descriptive text in t5xxl and a short summary in clip_l? Or do you keep clip_l empty? Or do you just say fuck it and write the same long description in both fields?
bruh, this is fucking insanewith a shitty 3060slow as fuck, surebut the quality is bananaswith a 48-64-128Gb GPUs the sky is the limit
>>101748164If comfy can run flux on cpu only, then would you be interested?
>>101748214ya with these we also will be able to use the promised animation model that Black Forest is working on, can't wait
>>101748164
>>101748226>promised animation model that Black Forest is working onO_O
>>101748182i'm a poorfag with no money for GPU upgrade, I have 64GB of system ram but only 6 on my graphics card
>>101748213Literally just describe the image exactly.>A still frame from a fantasy action movie of A snow warrior. She is wearing dark wolf furs, and a wolf rests on her pauldron. She is strong and dangerous. She is holding a dagger, and stalking her prey. Her eyes are yellow. She has a small amount of dirt on her face.
Prompting is gonna make me a better writer
>>101748213>>101748247Also I just combined both clips with a single string, so I am essentially putting the prompt into both t5xxl and clip_l
>>101748240https://blackforestlabs.ai/up-next/>no demo>open-source is, at best, only implied>no whitepaperUntil we get more than just a few cheesy marketing hooks, I'm not convinced it's not just hype to rope in investors
>>101748281
>>101748326too thick
>>101748240this >>101748301hype at its best .. well they delivered with FLUX so I give em some slack for now
>>101748301wewI mean, sora exists, so it's possibleI was a bit deflated after SDXL, but it's clear that we're on a specs plateauwe just need MORE POWERwith 128/256Gb VRAM, good quality video seems feasible, and the images are going to be insanely goodflux is 12b, then 64b-128b is where things starts to get indistinguishable from reality
>>101748329My bad g
>>101748353you're just pulling numbers out of your ass
>>101748353You are working on the assumption that more always = better. You can oversaturate it. Like over cooking a LoRA
I don't know why, but I never win at this game
>>101748356yes, but they're reasonable numberswe're here with 12Gb-24Gb VRAM /12B models, we need way more than that>>101748383yes, because that seems to be the case, more parameters = better models, = the need for more VRAMeverything about AI seems very much linearly scalable, at least for the time being
lllyasviel updating Forge again
>>101748403Does it though? SD3 has more parameters than SDXL and it is dog shit. SD2 has more paramaters than SD1.5 and it is dog shit too. There more to it than just making number bigger, clearly.
>>101748397maybe because you're holding it wrong
>>101748403>>101748430a big brain doesn't do shit if the data you feed it is shit
Flux does some kind of face replace and muscle tone thing.
>>101748434They just can't seem to figure it out. Model limitation I guess.
>>101748430haven't tested SD3 but they clearly fucked up somethingSDXL rapes 1.5 when it comes to realism
>>101748475the version of SD3 they gave us is seriously broken and is not even bigger in parameter size than SDXL, the architecture of SD3 supports big models of 8b, but they gave us SD3 medium with 2b parameters and it so gutted that it fails to make simple human anatomy .. they went so far to apologize>https://the-decoder.com/stability-ai-apologizes-for-disappointing-stable-diffusion-3-promises-much-improved-model-soon/fail at its best
>>101748475>haven't tested SD3it's a bit broken
>>101748453perhaps"holding the backs of playing cards""playing cards with their backs to the viewer""reverse playing cards""face-down playing cards"
Same as before, flux pro and other models:https://grain-compare-packard-anime.trycloudflare.com/
>>101746316Yes de3 is still betterWhat do you mean by prompt upscales, the rewriting de3 does? I've found that schnell has better prompt adherence than dev when given rewritten prompts that de3 spits out, but de3 is still better at fully vibing with the prompt even when forced to not rewrite, flux is sterile in comparison
>>101748545What was the consensus on this? Was it a legit case of dataset poisoning to censor and lobotomize this shit for (((safety)))?
>>101748584SAI is a money laundering company and making AI models is just a front they are not very good at.
>>101748584probably, men can lie on grass in SD3, women can't, go figure
>>101748611100+ million they grabbed, well they incubated the team that made SD1 into Black Forest .. so thank god they existed .. who long still? not that long is my guess
>>101748551Yeah tried those, they always show the card faces. >>101748568DallE3 is only better because it knows art styles better. LoRAs will be our saving grace if they can be trained cheaply.>>101748584100% poisoned for "safety". Can't make a woman laying down. Can't draw a nipple. Can't even draw a belly button. Still fucks up hands. Prompt coherency is fucked, it barely listens, you always have to fight against the model to get something good. Unlike flux, which just gives you what you tell it for the most part. I eventually managed to fenangle it but it was such a chore that it became completely un-fun.
>>101747917Breasts sizes are as vague as possible with the model, probably yet another thing that was scrambled.
>>101748542>>101748545lmao, I did see the news and a few videos about the drama, absolute shitshowwhat flux proves, tho, is that the trend seems to be that the more training data, (properly processed, of course), the better the results, but the more processing power it requiresthe good news is that, the current structure seems able to handle the load, and by that i mean that we don't seem to depend on any major breakthrough but just finding ways to scale the system we already have, which are probably very wasteful, and will also get optimized along the way
>>101747822Yes and yes.CUDA is an insane advantage for Nvidia, mainly because they started way earlier than everyone else.
>>101748368>>101748480>>101748661dope
>>101748542Don't forget the wonderful meme of "2B is all you need">>101748545Meanwhile on Flux
Some of the prompts I struggled with on SD3 I ran through flux and got what I was going for on the first gen every time.
>>101748584>censor and lobotomize this shit for (((safety)))Yes, and they didn't have enough talent to make up for it with something else like flux writing skills.
>>101748770Yeah because all the talent went and created Flux. Flux is literally the people who invented latent diffusion.
>>101746316Yes, DALL-E 3 is better because its dataset includes all the stuff lobotomized from flux.It's just a different approach : - Flux makes the model retarded for nudity, known people etc, which has the pro of not needing a moderation endpoint, but results in a worse experience than it could have been even with sfw stuff.- DALLE understands nudity, knows what porn is, artists, known people, but rewrites your prompts to not show them + relies on a second model to analyse outputs and block them.
>>101748811Anon, but you're wrong about Flux - it can still generate gore and nudity, and it does need a moderation endpoint which Replicate and fal do have.
>>101748834Generate a pussy with flux then. We wait.
>>101748811Oh, and also, I have access to dalle azure endpoints that don't check generated images for NSFW, and I can tell you that DALL-E 3 doesn't really know genitals (very, very badly, just a tiny bit better than Flux). And you can force it to use your prompt directly in most cases, unless it contains nsfw stuff.
Sometimes I have genuine difficulty discerning if these images are just screenshots or actual AI. I gotta look close to find flaws.
>>101748811>DALLE understands nudity, knows what porn is, artists, known people, but rewrites your prompts to not show them + relies on a second model to analyse outputs and block them.not really, they have a secondary system running that checks the output and gives you doggo when it appears in the gens, the model itself doesn't understand it
>>101748843It's enough that Flux can generate nipples and gore, it absolutely does need a moderation endpoint.
>>101748811>>101748843You are confusing flux pro with flux dev.
>>101748863The dog is only relevant to bing creator/ms designer, dalle does know tons of styles and characters.
can someone point me to the smaller flux files to dl for the 10-12gb range, fp8 i guess? i got a normal version setup but it uses 15 out of my 16gb vram even with lowvram
>>101748863>>101748811ah yea you said that.. I should finish reading before posting
>>101748834What I mean by "make it retarded" isn't that all images disappeared from the dataset, though they probably started with as few nsfw images as possible, but that the concepts are scrambled. Nude = underwear and all that.
>>101748084that would show in the diffusers code which you can look atat the end of the day you have a noise map that is the seed, you have the steps, that's the amount of noise that needs to be removed (n/steps), the model is rewarded based on the quality of the image at the end of the generationanything that happens during n/steps is what the model learned to achieve the final image, that means there will be some parts of the image changed over timethat is not to say they did not reward the model additionally for genital free generations which is why it likes to always put underwear on, you can tell SD3 did something like this because it prefers nude colored underwear
Do you all have dedicated PCs for this shit? Like a server specifically for inferencing?
>>101748882Just use the normal model and change your weight dtype to e3 or e4
>>101748876ya sure.. same tech tho.. right? anyhow if dalle3 was released they had serious problems, cause the dataset definitly knows more than it should>>101748551also playing cards face down are a real problem.. did not manage them shown back side yet
>>101748847I've had access to a less censored version for a while, and while it also still not that good for genital (better than flux, dicks exist for example), it's so much better for nsfw (lingerie, porn poses, etc).Too bad Dalle without moderation endpoints are the rarest thing right now.
>>101748863>relies on a second model to analyse outputs and block them
>>101748864>Flux can generate nipplesnot really, they look like candy at best>and gorenot really
>>101748904What? This runs on consumer hardware. I just use my gaming rig.
>>101748905i'm hoping to do imagegen with text gen so the smaller the better, i gotta reserve some space for the text model
>>101748904big boy PCs are 4090s with 64-128GB of RAM
>>101748904Running on my older am4 machine yeah, it avoids fucking up my main rig when it uses the available 24GB.
>>101748924>lingerie, porn poses, etcWhen we get controlnets and LoRA Flux will shit on everything else.
I have an rtx 3080ti 16gb. Should i dare to run flux and if so what ui should i use?I'm illiterate when it comes to ai
>>101748906>anyhow if dalle3 was released they had serious problems, cause the dataset definitly knows more than it shouldIt'd be heaven.NAI leak times 10.
>>101748937Try schnell for gore, for example. And nipples sometimes do look acceptable. Anyway, you're moving the goalposts, the point still stands: Flux *needs* an NSFW checker for user facing applications, it's not a 100% "safe" model.
>>101748948Comfy make an FP8 mode that you run just like a normal checkpoint, and included a sample workflowhttps://comfyanonymous.github.io/ComfyUI_examples/flux/#simple-to-use-fp8-checkpoint-version
>>101748959>Should i dare to run fluxYou can run it in FP8 (basically unnoticeable quality loss)>what ui should i use?Comfy UI is the only UI that currently supports Flux>I'm illiterate when it comes to aiEveryone is, even the engineers.
>>101748960Make that a 1000x, dalle3 leak would be a huge thing
>>101748882https://huggingface.co/Comfy-Orgfp8 versions no account neededThese contain the vae as well, which is why they're a little bigger than other fp8s
>>101748971What is fp8?
>>101748904ppl run this on as low as mid tier GPUs>>101748932ya ya, its to warm, I didn't read .. peace>>101748960it would be, but unlike NAI someone would need to write inference for it (probably would take an autist a day or two) and you would have MS and OpenAI on your ass, not a garage corp like NAI .. I wonder if the one who stole the NAI model ever got legal repercussions
>>101748972Yeah, the biggest jump in local ever, equivalent to a recent claude model leaking for LLMs.Will not happen, they keep that as safe as they can, but a man can dream...
>>101748882Whats the matter if its consuming 15gb?You have 16 for a reason
>>101748971>basically unnoticeable quality lossFor comparison sakeFP8, FP16, Pro
>>101748992>you would have MS and OpenAI on your assOnce it's out it's out, even if people lay low for months and share this on p2p only, they won't be able to do fuck all.
>>101748986floating point 8 .. the weights in a model are stored in floating point numbers of a certain bit length, the default for flux is 16bit precision, which makes the model 24gb in size, which will barely fit on a 3090/4090 (there is some overhead, so its less than 24gb) .. but reducing the length artificially to 8bit you reduce the size of the model (~12GB) and can fit it in less VRAM, but lose some quality (there were comparisons, its not much of a loss)
>>101748967literally me having never used it>>101748981>its even single filesweet, thanks for the spoon feed anons i appreciate it. 16gb isnt much to play with for vram but a text model is going to need at least some space else it'll be too slow
>>101749013The reflection effects on pro are good.
>>101749013this >>101749020someone steal PRO pls
>>101749020I know, pro is fucking sick, so I don't think they need to be worried about community fine tunes on Dev.
>>101749016Holy shit i just bought this graphics card and i'm already a vramlet
>>101749007i want to be running an rp at the same time but occasionally inject images of what it thinks the surroundings are (based on my lorebooks and stuff). i can adjust the amount of layers the chat model uses, but to gen an image it'll just crash if it doesn't have the amount of vram it wants. sd already works fine, but i want to plug flux in next
>>101749048get a second gpu for rp
>>101748986fp8bp
>>101748561Added more configuration to pro from replicate - step count and interval (pro-only feature, apparently influences variance by the cost of prompt following, idk how its different from guidance)
>>101749060you need like 6 of them for this hobby. at least i can shift around some things and run both at the same time (hopefully with these smaller files) like i do with sd
>>101749048Thats nice. What llm are you using?
>>101749048In that case use schnell, it gens in 4 steps so your gpu will be whored less. 4080 and 32GB RAM for the spill over and my ram is sitting at 26GB used. GPU 100%. >>101749060This, pretty much. It's a phat model.
Anyone has a workflow for tensorRT for flux?
Flux is the nazi chud model
>>101749138prompt?
>>101749122Damn right.
>>101748243I don't get it. That's enough to run it already.
>>101749161>This image is a high quality 2011 screenshot from a TV anime which is visually striking for its rich pastel color pallet soft line art and three tone shading creating a soft and inviting visual design. The image shows the two main characters in a locker room, one of the girls is pulling up a pair of tiny pink cute panties over her wide hips and plump bottom causing them to ride up her butt crack slightly, the other is pulling her shirt over her head, undressing, The view of the scene is partially obscured by black lines giving the impression that the camera is behind a grill or vent. The subtitles read, "I think I heard something from that locker."
>>101749209thank you, appriciate it.
>>101749208how long per 768x on 6gb?
>>101749091l2 miqu has been my main for a while, but command-r+ 103b and the new mistral large 123b are very good without any tuning. i've been spending the most time lately with the new llama 3.1 70b instruct and non-instruct but i haven't made up my mind about it yet, sometimes i think its not any smarter than l2 was and writes worse for rp.>In that case use schnellnow that i saw that name i remember reading thats the smaller one. good call. i'm not even looking for characters, just broader stuff like a forest in winter, to go along with the scene. so i think i'd be fine with whatever the literal worst version of flux is as long as i can keep the vram usage down
>>101749230Schnell isn't smaller, it's the same size, but it will gen an image in 4 steps as opposed to 20 on dev, and they will be good enough for your purposes I think, but not as good a dev.
>>101749229100-120 secs
>>1017492564 steps just isn't enough for text and fine details, the model has to infer/assume too much per step.
>>101749122lets change that then
>>101749257That's actually pretty decent. How much RAM? 32GB I assume?>>101749261I don't think hes embedding text for his images, he is generating scenes from his rp. Text doesn't seem like a requirement, nor does fine detail.
prompt: emerald cheese
>>101749290tasty! I want it
Took about 12 minutes
>>10174928516GB. It's limited to fp8 and Schnell model though. fp16 will rape my ssd pagefile drive, and dev model you will need more patience, around 20 minutes for 20 steps of 1024x gen.
>>101749256why is it 4 steps specifically? steps was always an option before wasnt it? (been out of imggen for over a year)
>>101749122>Flux is the nazi chud model
>>101747237>https://files.catbox.moe/clnq7v.pnghonestly i thought you were retarded but now that im trying it myself i think you're on to something, it naturally gravitates towards clothed when you put up the step count even though there is no mention of clothes
>>101749339It's just the way schnell was made.It's able to output coherent stuff with just 4 stepsNo idea how they did it.
damnit flux, DEATH is card 13 .. not 12
>>101749342It gravitates towards clothes because nudity is extremely limited in the dataset, I would bet that nudity is counted in the thousands in the dataset
>>101749377and their caption model probably avoids describing the nudity, you see it happen with open caption models like CogVLM
>>101748355
>>101749348You have a teacher model (50 steps) and a student model (4 steps), you train the student model against the outputs of the teach model.
>>101749398I mean there are some tests you can do when the dataset isn't properly captioned, for example, truly uncensored but poorly captioned will show you genitals if you get the right angles (like butt from behind and below on a woman). But when you do that you will see a thong/underwear.
>>101749328>around 20 minutes for 20 steps of 1024x genThis is how long it takes me on a 4080 with 32GB RAM of I run in fp16.>>101749339It's a "turbo" model. SD has one too. SD has a cascade model which does near real time generation. We have lots of toys since last year, multiple competing models too, and all of them have been made obsolete by Flux.>>101749342>>101749377This is 50 steps.https://files.catbox.moe/stxrq5.png>>101749398It's probably closer to this. Like how copilot sees blurred faces.
>>1017493484 is the minimum steps with decent visual result. Though the manual recommends to use 4-8 steps but you still can use more steps apparently.
>>101749400Oh, clever.
>>101749440it's why it's a one-way trip though because you essentially kill the student model for learning new knowledge, so to train the student you have to retrain it back to 50 step because high steps is a crucial part of teaching diffusion models
training cunny lora>43.15it/swish me luck bros
>>101749437i still dont know why this is different than dev though. is 'steps' not the same as in auto1111 selecting 30 vs 40 steps? is schnell the same size in vram/ram as dev or smaller?
>>101749472staber guy strikes again
>>101749475A model works by denoising n/stepsSchnell is trained to remove 25% noise per stepDev is trained to remove 2-5% noise per stepThat's the difference
>>101749475schnell was forced to learn how to generate the same images as dev in fewer steps, this necessarily causes it to have issues with some things since each step is a bigger change to the image.
>>101749475>schnell the same size in vram/ram as dev or smaller?same size>>101749475>is 'steps' not the same as in auto1111 selecting 30 vs 40 steps?yea, but depending on sampler and model type more steps doesnt mean better output, if the sampler isnt something like ancestral it just will converge to an output at certainpoint and more steps more or less just shifts a few pixels
Btw anons, please don't forget that: flux.1 pro is the original flux.1 dev is a normal distill from proflux.1 schnell is a 4-step distill from pro (not from dev, obviously)
>>101749321You bent your GPU pins!>>101749437I was running 20 steps constantly before I realised what schnell was. I thought it was the smaller model, didn't realise it was a turbo model. So all my first gens are on schnell, and I was happy with the results, but didn't realise why I was struggling so hard with the text. >>101749472>43.15it/s>36 hoursOn runpod this will cost about $50. Not sure I am willing to pay that just yet. Interested in your results, even though I hate cunny.
>>101749513pro leak when?
>>101749489so the dev one just does more work per step/pass. why cant they just say that lel
>>101749526they do, if you know how to read
In the next few years it'll be a miracle if we get any woman at all in the released base models, clothed or not.
If you want some non-autosaging bread then it's right here...>>101749563>>101749563>>101749563
>>101749513pro is just dev + secret prompt rewriting (oh i mean, 'interval') to farm $$$dev is just a poorly optimized handicapped version to make schnell look goodschnell is the real one
>>101749604Yamada is that you?
>>101749395prompt please
>>101749603
>>101749628>emerald berries
>>101747863did you notice any difference when trying different values?
>>101748584>Was it a legit case of dataset poisoning to censor and lobotomize this shit for (((safety)))?yeah it's completely a lobotomy, they probably removed all nude people in the dataset so the model has no idea how anatomy works, I'm glad flux exist, SAI can die now for all I care, they deserved their downfall
>>101749341that hand on that bird is so uncanny kek
>>101749321How many steps?