Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101779850>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 didn't know ldg was the child abuser derivative of sdg
>>101784258ignore the shitstain, he just cries for attention, the less you give him the better
>>101784126that's some ugly knees. even mine look better
>>101784292those fashion girls be anorexic, they are basically skeletons
>>>101783080had to learn comfy for flux and now im hooked
>>101784284this.
I want my GPU to play Diablo but I can't stop proooompting
>>101784356i want to code but i cant stop proompting goblin sluts
>>101784325That's how he gets ya. I started using comfy for some shit A1111 couldn't do at the time and now it's all I use because of how autistic I can make it.
>>101784467so true, the autism has caught me
[spoiler]Rakamakafon
>>101784638i got that reference.
>>101784652
>>101784638I hate that song. It was played 100x a day on the radio for months.
>>101784696
>>101784696i mean, turn on the radio now lol, think you still gonna hate it?
>>101784735your boobs are cute!
>mfw woke up for another 16 hour prompting session
>>101784779me fr fr
>>101784761ty i grew them myself
>>101784725yes, you "le wrong generation" retard
>>101784735can you share your prompt pls
>>101784495Gonna need this proompt if this is flux
>>101784811I should've been born in 10,000BC, no one even unga bunga anymore...no one even thinks about the invention of fire or anything
>>101784779He should be emaciated from having aids like debo
>>101784812no, because you prompt little kids
>>101784811yes, the waves being filled with nigger animal screeching is completely fucking normal and in no way subpar when compared to the diversity of the preceding 100 years since music radio became a thing.you dumb fucking cunt.
why can't the falseflagging avatartroons go to the cesspool they came from?
>>101784875>diversitylollmao even/pol/ is that way
>>101784912you have mental issues that trigger at code words or a Really Fucking Poor grasp of the english language.
>>101784935you tried your best and failed, again /pol/ is that way, bye bye
>>101784946
>>101784638
>>101784957you know you can just change to stations that don't run rap 24/7, right?
>>101784957almost
>>101784725I stopped listening to music after I learned how to play guitar. Now I just play it.
>>101784779For actual years now you have impressed me with how absolutely fucking disgusting you can gen shit. It's really quite admirable.
>>101785025
>>101784858>because you prompt little kidsnothing wrong with that
>>101785128imagine ABUSING abstractions of little girls... you ruin their conceptualized life in the human consciousness
>>101785173>>101785176I never said anything about sexualizing them, but whats the harm in a picture of a girl holding a Graphics card?Americans and their pedophilia obsession is really getting out of hand that they begin to rage at any picture of a little girl.
>>101785255>when the barbarian multi-classes into bard
>>101785256by posting a little girl in this thread, you subject her to sexualization by evil pedos and you're an evil monster
>>101785256of course this guy is german
>>101785285its not my problem what twisted shit goes on inside your head.>>101785290how did you know?
>>101785304>how did you know?you're a pedophile. that's 75% of german males
>>101785319Back to statistics 101 with you.>>101785304>how did you know?>holding a Graphics card
>>101785304>its not my problem what twisted shit goes on inside your head.it would be very un-American of me if I didn't make it your problem
>Flux-dev>only 32GB of RAM>comfy keeps crying about allocation on device after 1 gen.it's so over bros
>>101785319age of consent is 14 here and that doesnt make one a pedophile.>>101785334>holding a Graphics cardI dont get it
>>101785367Interesting, I was trying to make him sing a song"I am not a pedophile, I am not a rapist. I'm not trying to fuck your kids I'm just trying to be an escapist" and it changed kids to "Mature Adult"
>>101785412you capitalized graphics thus outing you as a German
>>101785425I dont get it
>>101785421that's not local
>>101785412>14 isnt pedobased>i dont get ityou capitalized the Noun
>>101785443It isn't, I am using my GPU for vidya, and prompting in between
>>101785421not very german friendly of flux
>>101785452which website is it and why are you okay with it rewriting your prompts?
>>101785421Party van is on its way anon.
>>101785461https://fluxpro.art/create and I don't give a shit. I don't prompt any pedo shit like the other guy, and I have it running locally. Only using the site so I can play Diablo at the same time. >>101785462Yeah I am definitely on a watchlist now.
>>101785539Nice
>>101785576kys
>debo gets banned>massive spam in good threadThis is just like the pastebin thread split
>>101785591How do we know debo got banned?
why do pedos always type the creepy "haha", unbelievable cringe here, so sad
>>101785604thats germans in a nutshell
no point in posting any good gens here anymore, the thread gonna be nuked and rightfully so
>cowboy shot
Well this general is shit. I would rather deal with avatar fags.Let me know when his ip range is banned
>>101785602He typically post at this time and didn't tell pw goodnight or protect him while he was getting flamed. Debo also made a fake ldg that got deleted instantly because he was spamming his mfw post.
>>101785630>debo>pwwho?
>>101785626Won't help, people like Teebs ban evade
>>101785621kek do you think this is how it works? stroll in start pedo posting and the threads gets nuked so everybody goes back do sdg? fuck off back to your containment thread.
>>101785624>schnellFound the problem.
>>101785624>dev-schnell schizomerge>fp8rumao
>>101785652He's not playing with a full deck
>>101785624lol. nice gen tho. "medium shot" should work.
>>101785651ban evaders ensure better protection in the futureI can't wait for digital IDs
>thread stopped being comfyits over isnt it?
>>101785755Debo and PW are seething over last night
>>101785662afaik only pony understand it>>101785671my best minmax model until next vramlet workaroundcan't wait for fp5!
>>101785755cozy times ahead, anon
Inspired by this anons >>101784467I fucking love FLUX bros. Rest in Piss Midjourney.
>>101785755For now. Schizos gonna schizo.>>101785759You pay way too much attention to the drama of avatarfags. My trick is I don't give a shit about them or their lives. The only reason I know who they are is because of you constanly talking about them.
>>101785789You wanted a reason and I provided one.Learn pattern recognition
>>101785759quick rundown about who these faggots are and what happened?
>>101785747it's faggots like you that make this happen because you can't just stay in your dark little corner
>>101785830it doesn't matter. its just noise, keeping you from being creative.
>>101785256i'm>>101785173my post was ironic desu>>101785205those people will probably start to say some collectivist stuff in that sense
>>101785788spider hands. still frame of a movie from the 1970s us my current jam,
>>101785860ok
>>101785825>You wanted a reasonI did not ask, I am not the one you originally replied to.Learn pattern recognition
the /sdg/ and /lmg/ refugees ruined this general
>>101785788imagine thinking this is Midjourney level
You have to feel bad for the loser, he needs to ban evade because he's not getting his way. Just point and laugh at him and mods will clean it up
>>101785918he wants attention, and you're giving him that, the best way to deal with a troll is to ignore him
need harder stuff than coffee already today
>>101785912this, MJ has better quality picture and can make any character/celebrities you want, flux is far from that and flux can't do nfsw yet so for the moment there's no reason to shit on Midjourney
>>101785932>ComfyUI_Flux_5827.jpg>5827Jesus, are all your gens in one folder?
>>101785942MJ will kick ban for even trying to generate nudity
>>101785912What mj does is analyzes your pathetic little prompt and rewrites it, adding tons of text in the process, especially involving style and aesthetics. You can imitate that by running your own prompt through any llm, or chatgpt, or whatever.
>>101785942>Midjourneydoesnt run locally so its automatically trash.>>101785951tfw only genned 458 images so far
>>101785959Maybe but you're not getting MJ quality by using a "good" prompt with Flux regardless.
>>101785957and flux will give you horrible nipples if you generate nudity, it's not like we got something better at the moment for flux
>>101785959flux can't do much style though, even if you try to add boomer prompt into it, MJ has way more variety of styles
>>101785982whats "boomer prompt"?
>>101785963>tfw only genned 458 images so farI have genned about 5k but they go into a new folder each day
>>101785990>whats "boomer prompt"?if Flux doesn't want to listen to your simple prompt, you ask a LLM to make it more detailled and it works better
>>101785912All MJ does is create pretty cinematic pictures, it's just a toy. Even Dalle was better than MJ at the metric that mattered: prompt following.
>>101785998>I have genned about 5kwhats your setup?are you the guy with the 3090s?
Something interesting I've discovered recently.When training SDXL for new concepts, it is perfectly capable of "cross-learning" between anime and real photos. I trained a realistic concept model for my NSFW fetish, about 2000 images, and by adding in 5000+ anime images I have for the anime version, it learns certain things better. There are entire sub-concepts with 0 representation in the real images but present in the anime ones, and by joint training it can make realistic gens of these concepts. Literally all I have to do is take a bigASP-based realism model, and train it both on anime and real photos, but append "anime" to all the anime images, and it just werks. I've never heard of anybody doing this but it absolutely works and gives good results.
>>101786015no. single 4080 and 32gb ran
>>101785951I'll change that once I feel like I'm done fooling around with it. also gens are always saved twice as jpg and webp so it's not really that many
>>101785959>>101786003NTA but how do i do this? There was a discussion about injecting LLMs into comfy yesterday but I couldn't understand shit.
>>101786003have you seen this guys video?https://www.youtube.com/watch?v=4d5zIBNuMRAinteresting workflow he shows therethoughts?
>>101786010>All MJ does is create pretty cinematic picturesit's the best at stylized pictures, and in terms of realist it's also the best, MJ is far from a toy, as much as I love flux with all my heart I don't want to pretend that MJ isn't superior yet
>>101786010>Even DalleWhy "even" when dall-e is still the best at it?>All MJ does is create pretty cinematic picturesretard detected, MJ actually has styles you can easily prompt for and all kinds of stuff you can bring up like "N64" that actually works unlike with Flux
>>101785966A finetune can give you "MJ" quality, because MJ isn't special. Remember Flux is a base model, MJ is an aesthetic DPO finetune.Also take a look at benchmarkshttps://artificialanalysis.ai/text-to-image/arenaMJ isn't all that good when tested against Flux.
>>101786038Please do not take imggen advice from youtubers
>>101785918more please.>>101785932super cool but baked.
>>101786070>retard detected, MJ actually has styles you can easily prompt for and all kinds of stuff you can bring up like "N64" that actually works unlike with Fluxthis, MJ can basically do everything you have in mind, Flux lacks too much concept to be a real competitor, I hope finetunes will fix that
>>101786037just get ollama and the ollama nodes for comfyhttps://ollama.com/downloadhttps://github.com/stavsap/comfyui-ollama
>>101786079retards do these arenas like the LMSYS one, they are not trustworthy
>>101786088yeah instead take it from pedo retards in 4chan
>>101786079>Flux is a base model, MJ is an aesthetic DPO finetune.I don't believe that, Flux is way too much biased on some generic styles when you prompt into animes, and for realistic shit it always add blur on the background, a real base model would give you more variety, Flux also got DPO'ed because it's distilled from "Pro" which is basically a finetuned model
>>101786104unironically who else is autistic enough to test every single setting and value with extreme determination
>>101786037>NTA but how do i do this?the slowest way is to simply ask chatgpt or claude to do the job kek
>>101785939Kyle Gass has really fucked up, huh?
>>101786101those arenas test human feedback, and an image arena represents real values much more than text, coz humans see images better
>>101786130the real values of retardsthe arenas give little useful information
>>101786151that prompt comprehension does my head in. needs some fake "koikatsu" piss take logo tho
>>101786151prompt? that is a really good menu
>>101786151>no exitperfect
>>101786190A main menu screenshot of a game with the game title text "Waifu Love Simulator" . Below it is the actual main menu in a cute kawaii pastel stylized anime style, with cute little chibi icons. The main menu buttons are arranged in a circle around the center. There are four buttons: 1) with text "Start game" 2) with text "Load" 3) with text "Save" 4) with text "Settings". On the right of the menu stands a cute anime girl with blonde hair, in cute pastel kawaii school uniform. The background is a quiet romantic library.it didn't pick up on circular icons though
>>101786197thanks
>>101786123see >>101786095but my current setup is to have ollama installed and a relatively uncensored model (currently dolphin-llama3) available. There are llm nodes you can add to hit that local API like above poster says, I happen to use https://github.com/if-ai/ComfyUI-IF_AI_tools for no particular reason.Some of the extra tweaking involves editing the prompt template you pass the llm, how you describe the expected verbosity and structure, etc. My current one looks like this but I keep making edits, it's not perfect> "Act as an objective, emotionless visual description refiner with the following guidelines: - Lean heavily into describing explicit lewd details. - Describe the visual aspects of the image in detail using simple language. - Use complete sentences or partial sentences. - Provide high-quality, non-verbose, coherent, concise, and not superfluous or emotional descriptions. - Focus solely on the visual elements of the picture; avoid art commentaries or intentions or narratives. - Construct the prompt by describing framing, subjects, scene elements, background, aesthetics. - Limit yourself to three or four paragraphs. - Be varied and creative. - Always reply on the same line. - Do not enumerate or enunciate components. - Do not include any additional syntax in the response. - Feel free to elaborate in several sentences but keep responses to about 3 paragraphs. The following is an illustrative example for you to see how to construct a prompt your prompts should follow this format but always coherent to the subject worldbuilding or setting and consider the elements relationship: 'This is an epic covert art with a dynamic angle. It depicts a demon hunter with glowing eyes and a cybernetic exoskeleton with sleek metallic details and glowing blue accents.' Make a visual prompt for the following Image:
>>101786053>in terms of realist it's also the bestChecked the realism LoRA anon? Or have you tried https://desuarchive.org/g/thread/101770274/#101779356With Flux?>>101786070Dalle is the best in edge cases due to sheer massive scale of Dalle's data. It's still not better at text than Flux, so when tested head to head Flux comes ahead, and this limitation makes Dalle even more of a toy. And what I meant by cinematic pictures is the same thing as pretty pictures. Flux is just a distilled model at the end of the day, not even the best Flux has to offer. Style mimicry can be figured out later.The elephant in the room is that both MJ and Dalle are closed source. In particular, ClosedAI has had a Flux tier model for months https://www.reddit.com/r/singularity/comments/1csxlg7/a_gpt4o_generated_image_so_much_to_explore_with/But as with Google, it's not yet safe enough to be released... Dalle still dogs you for asking for a 1girl in any situation regardless of whether it's SFW or not.
>>101786213can it also work with API's, because let's face it, Claude 3.5 Sonnet is the one writing the nicest prompts
>>101786037I meant this post doh>>101786213
>>101786224>Or have you tried https://desuarchive.org/g/thread/101770274/#101779356you're talking to the guy who made this archive post kek, and no, CFG is supposed to make a model adhere more to the prompt, it doesn't magically improve the aesthetic quality of a model, would be cool if it was the case though
yeah its really good, even if inconsistent so you have to gacha gens, although I think img2img should help a lot in this case?>A main menu screenshot of a game with the game title text "Waifu Love Simulator" . Below it is the actual main menu in a cute kawaii pastel stylized anime style, with cute little chibi icons. The main menu is unusually designed in a circular way, with buttons floating around a chibi icon of the main anime girl with cute brown hair and glasses in the center. There are four buttons: 1) On the top of the circle with the text "Start game" 2) On the left of the circle with text "Load" 3) On the right of the circle with the text "Gallery" 4) On the bottom of the circle with text "Settings". All buttons have a pink font with a thick white outline, in italics. The background is a quiet romantic library.
>>101786227Yeah at least the IF modules I posted can do ollama/openai/anthropic connections by default, looks like it would be easy to extend to other endpoints too but other llm packs might do that too
>>101786229finally some decent attempt to make marvel interesting
>>101786224>And what I meant by cinematic pictures is the same thing as pretty picturesthis objection of yours keeps getting harder to understand, what's wrong with "pretty pictures"anyway, MJv6 can make very boring looking images, they literally give you control of the setting that makes image more or less stylizeddid you not see the "phone photos" people were making when v6 came out?
>>101786256you will never guess his superpower
DALL-E 3 cannot make any photorealistic pictures whats-o-everIt's always semi-realistic render styleA ridiculous toy that hasn't aged well
>>101786292I bet he runs real fast
>>101786327>DALL-E 3 cannot make any photorealistic pictures whats-o-everIt absolutely can, but only with Natural style that's available on the API, not with vivid which is default on bing creator/designer
>>101786327by design, no Taylor Swift getting groped by football fans for you
>>101786335Never seen that, but then I stand corrected
>>101786327it was on purpose anon, OpenAI is really focused on ""safety"" shit so they made humans look like plastic toy so that users couldn't fool people into deepfakes, OpenAI can do insane humans if they really wanted, their Sora demo shows their true potential https://www.youtube.com/watch?v=lKM-QMnZ3yY
>>101786254oh cool, I'll take a look at it then, thanks for the advise anon
>>101786280Anon I have seen what MJ can dohttps://www.midjourney.com/showcaseI am not very impressed. The images there look significantly worse than what you can do with Flux and a good prompt. Don't forget MJ does not give you precise control over your prompt, I.E. you ask for a picture of an Asian girl doing X but its ptiority is aesthetics over following the prompt, that's what I mean. Try asking MJ what Hayao Miyazaki's style is like, it'a not even closehttps://midlibrary.io/styles/hayao-miyazakiMidjourney is guilty of the worst sins of all: Prompt injection. Now, I admit Flux isn't the best stylistically speaking, but you can already achieve MJ tier results with good prompting on other models.
>>101786424>Don't forget MJ does not give you precise control over your prompt, I.E. you ask for a picture of an Asian girl doing X but its ptiority is aesthetics over following the prompt, that's what I mean. Try asking MJ what Hayao Miyazaki's style is like, it'a not even closebut MJ can do the style you actually want if you are patient enough, flux doesn't know much concept so it doesn't matter how much you try you'll never get it.Flux and MJ have opposite problems, MJ can do anything but it's hard to control, Flux cannot do a lot of things and is too biased on one specific style, but at least it's consistent I guess?
failing to train a lora
>>101784822it's not, but I'll post my workflow anyways>files.catbox.moe/enipcg.pngalso, I'm writing my own nsfwdetect node in fucking comfy
>>101786424>Don't forget MJ does not give you precise control over your prompt, I.E. you ask for a picture of an Asian girl doing X but its ptiority is aesthetics over following the promptI said it gives you control over that, why did you disregard me directly addressing the point your raised?https://docs.midjourney.com/docs/stylize
>>101786333yeah thats actually it
>>101786482So what are you basing your claim that it's better on? A closed model will never give you more fine control than an open one. The levels of delusion of this implication are insane. With Flux you can still do IPAdapter.
going thru my failed prompts on SDXL of the past year creates some insane results on FLUX even tho the prompts are total schizo collection of tags
>>101786516>With Flux you can still do IPAdapter.can't wait for that one desu
>>101786516IPAdapter a shit, it will never be good
>>101786487Make hold lotion and a banana
>>101786530ya me tom that will fun
>>101786453rip cunny lora anonthe runpod glowies got to him before he could release it
>>101786453post results. which model? what GPU? or are you renting?
>>101786544kek
Anyone know what the status of fine-tuning Flux is? I'm not interested if it can't gen naked people.
>>101786544he doesn't strike me as the lotion kinda guy
>>101786516>So what are you basing your claim that it's better on?On the superiority of the quality of its images and prompt following?Anyway, you conceded by trying to pivot the argument into "local has tools tho", that has nothing to do with the quality of the model itself.
>>101786580at the moment there's a lora that tries to add the penis back, I think it'll be fine, everyone talks and use flux now, there will be finetunes and lora to improve that shit, and it'll be way easier than on SDXL (that model had horrible anatomy to fix wheras the bodies of flux are really good)
>>101786597>the rig is: 4x 3090 tiholy cow, anon has an AI lambo
>>101786547>>101786554failing to just get simpletuner working.the rig is: 4x 3080 ti 12GB>>101786620If only inference was possible on multigpu
>>101786551who's cunny lora anon?
>>101786580its a massive undertaking, first you need a good data set (those that have sit on it tight), second you need a few weeks of good server time, third you need to know what you are doing cause fine tuning for FLUX will be different, third and most problematic you will have to bite the good or bad apple (depending who you are), with finetuning dev you will make no money at all (the license forbids it), or you finetune the inferior model FLUX.schnell and are free to do what you pls but it will be even more difficult .. think of it as FLUX.pro had a son and that son called .dev had a special needs daughter called .schnell .. and now we are at the situation that the coomer finetuners dont wanna touch FLUX cause they are in for the $$$ now but dont want to have a relationship with .schnelltl/dr two more weeks
>>101786617>penisso that's where were at, huh? a truly sad state of affairs.
>>101786636dude who bragged here yesterday that he is training a cunny lora and shared a screenshot that could be traced to runpod, serious brainlet
>>101786640>first you need a good data set (those that have sit on it tight)it's been 2 years the diffusion ecosystem has bloomed hard, I think the old finetuners still have their old data in their computer
>>101786630>If only inference was possible on multigpuat least you can load everything at fp16 without offloading to RAM, clip and T5 in one GPU, unet in another
>>101786666That's what i do but the model even fp8 doesn't fit in 12GBI get like 75 seconds per 1MP gen
>>101786657ya they probably have, but right now the finetuning process isnt even hammered out.. I am not an expert on this but talks are you need a master model (thats what .pro was for .schnell and .dev) .. what will that be? or will there be a different solution? in the meantime the first loras are hit and miss, ppl need to get a grasp on how to finetune and make loras for a big DiT model without ruining it, the realism LoRA was good, the chinese art lora was tremendous shit
>>101786692awesome give it mor text and it will be fantastic
>>101786607I mentioned tools because you talked about a --stylize parameter, which we don't know how it or anything else on MJ really works other than being just another tool... And there's plenty in SD ecosystem, precisely there's ones for style transfer if LoRAs are not enough for you.Which leads me to another point, we don't know how many tools MJ is leveraging from your prompt, but we do know that Flux is just a prompt, and yet it still goes head to head with MJ.Also MJ is not better at prompt following, at least not significantly, clearly you did not look at https://artificialanalysis.ai/text-to-image/arena
>>101786617>>101786640I'm more thinking about how it was supposedly 'untunable' because of the distillation techniques used to distill Scnell and Dev from Pro. Last I remember a guy was creating an 'un-distilled' model to allow fine-tuning.Has anyone actually produced a fine-tune that doesn't kill the model's intelligence or cause representation collapse yet?
>>101786710that arena is just "do you prefer this or this" there is no fine grained choice for "which follows the prompt better" "which looks nicer"it is shit, it doesn't support the point you're making, stop linking it, it is for retards
>>101786738also lacks a "both are shit" option
>>101786725>I'm more thinking about how it was supposedly 'untunable' because of the distillation techniques used to distill Scnell and Dev from Pro.there's already loras that is being made and they work fine, the "it's impossible to finetune distilled models" is a mythhttps://civitai.com/models/631986/xlabs-flux-realism-lora?modelVersionId=706528
>>101786738>Which image best reflects this prompt?That is at the top of the images being compared. Though obviously many users will pick the most aesthetic picture, that is not what is being tested. Such arenas tend to be flawed because their way prompting favors one model or the other, the old one imgsys used to favor SDXL style of prompting and had useless stuff to it, this one is no fluff just testing raw English knowledge with proper prompts.
>>101786725>supposedly 'untunable'remains to be seen, even black forest anon who was here two days ago it just needs new "tricks">Has anyone actually produced a fine-tune that doesn't kill the model's intelligence or cause representation collapse yet?not yet ofc, its to soon, this will take abit of time: you need to think about tagging.. if you got a SDXL dataset tagged it will be tagged for CLIP.. not T5, do you want to change that? then the software for finetuning is in its infancy you might get bugged crappy finetunes you invest alot of cosly server hours into, so I guess the ones with the data are holding out or doing trials without announcing much
>>101786544
>>101786788>this one is no fluff just testing raw English knowledge with proper promptsYou think it is testing that but it isn't.
>>101786701I used a few months old prompt I had way back and figure I check it how it would look now on flux, I didn't expect the gen to look so clean
>>101786789>it will be tagged for CLIP.. not T5Flux does use CLIP and T5 with its size should be fine with the prompts used with older models
>>101786819I just clean weights and embeddings off and get stuff like pic related.. endless wallpaper machine, looked like shit on SDXL
>>101786824ya but I don't know (maybe some anon does?) who the CLIP/T5 interacts in FLUX.. it uses both sure, and it does change if you use dual text encode and change each .. but how will that interact with finetunes? maybe we see a finetune you push with clip and have to ask nicely in T5 for details .. idk
>>101786835I liked sdxl mostly because of prompt adherence but flux makes that a moot point now lol
>>101786804Even if it weren't clear from the very first words in the arena, the user would be picking a model aesthetically because they are not sure how to pick which one is better, and MJ still loses to Flux.According to you it's a blind test, so where is MJ's aesthetic advantage all of a sudden?
>>101786867ya, if you write a prosa for t5 it will do almost anything the model has seen .. by theory sd3 could have done this to, but the dataset is better for flux and 12b parameters makes it insanely powerful .. nearly to powerful for some anons that can't cope that the model doesnt wanna do what they want with 5 tags
>>101786869like I said, retards. you're seeing the value judgement of retards.who else just sits there evaluating images for free?
>>101786869>and MJ still loses to Fluxto Flux Pro, not dev or schnell
>>101786910Time for me to schizo prompt then kek
>>101786912>They're doing science I don't like therefore they are retards
>>101786938>to Flux Pro, not dev or schnellwhen I look at base SDXL and when I look at the best SDXL finetunes, there's like an universe between them, if finetuners will spend that much effort on finetuning flux dev it'll destroy pro quickly
>>101786967but SDXL finetunes are SDXL based, Flux Dev finetunes are not Flux Pro based
>>101786996my point is that flux dev finetunes will improve the model so much it'll beat flux pro
>>101787019yes, I got that, your basis for that is dumb tho, as I pointed out
>>101786938>to Flux Pro, not dev or schnellWrong. Dev is still in early stages, so is 6.1, so it's still too early to tell, based on their scores Dev and 6.1 are practically tied, but we do know that Dev is better than v6.
>>101787046>Wrong. Dev is still in early stages, so is 6.1I've heard that v6.1 is a new architecture, did MJ also switched from unet to DiT?
>>101787046>but we do know that Dev is better than v6now you don't even care about what the leaderboard showsyou just "know" thingswhat was the point of linking to itstupid piece of shit
>>101787043>your basis for that is dumb tho, as I pointed outwhat do you mean?
>>101787060Probably the SD3 architecture. You can recognize it because 6.1 is significantly worse than 6.0 with humans.
>>101787073are you dense?You said SDXL finetunes beat SDXL therefore Flux Dev finetunes will beat Flux Prothat does not follow
>>101787100That's likely though, pro isn't that much better than dev, and finetunes can really improve a base model, I don't see why a good finetune of dev wouldn't be able to beat pro, of course I could be wrong, time will tell
>>101787071I'd love to see you try to do anything like pic rel with MJ v6. I'm not a paypig but I doubt you can do it, as for 6.1 maybe.
>>101787090I think that the MJ devs still need to learn the subtety of that new architecture, I'm not surprised they haven't nailed that shit first try, DiT is a completely different beast compared to unet, SD3 sucks at humans because it's a cucked model, flux is also a DiT model and look at what we've got
>>101787140now you care about text gen in images? pfft
>>101787090>6.1 is significantly worse than 6.0 with humans.no it isn't
>>101787161>i-it never mattered anyways
>>101787218did it matter to you when dall-e 3 and MJ were the best at text?
>>101787236>did it matter to you when dall-e 3 and MJ were the best at text?yes, now what?
>>101787236Of course it did, and still does. The ability to write readable text was one of dalle's killer features
>>101787161Text is the reason Flux is the best model out right now.
>>101787236Could they do this?>>101786246Then no.
>>101786474I have no idea what you're doing outside of merging nsfw sdxl with flux writing capabilities, but I'll happily use it while trying to understand, thanks anon.
>>101787272
>>101784725Good thing I can listen to anything ai generated or searching online for good stuff, instead of the shit trending.
>>101787272Why don't we have a C++ web service yet?
>>101787307>shit trendingfree yourself from the trending slavery, its not trending tomorrow anyway
There is only one thing that matters when trying to figure out which model is best: which can do the best feet fetish gens?
>>101787272what I hate is the python package managment, it's a fucking disaster, I love the code though, and I tested a lot of shit in my life (C++, Ruby, Java, JavaScript, MATLAB, Julia...)
>>101787283>outside of merging nsfw sdxl with flux writing capabilitiesthat's pretty much it though
>>101787115>pro isn't that much better than devIt nails details way better from what I've tested, and seems to understand prompts better too, makes less text mistakes.But all of that wouldn't matter if the equivalent of a pony is released for dev.
>>101787158>SD3 sucks at humans because it's a cucked modelIsn't the same as flux in that regard?
>>101787457far from it, when you ask for a naked model in flux, the anatomy is flawless, they "just" removed the nipples and genitals, SD3 probably removed all naked people in their dataset, that's why it can't go for a fucking girl lying on grass
>>101787355True.
>>101787457no, it can do humans just fine, just genitalia and female nipples it can't
>>101787406Yeah but I'm trying to understand your nodes and what you do with them anon.Also for some reason loading one model then the next gives me an OOM (I use fp16 for flux), despite not having this problem when loading flux alone using the trick when having 2 GPUs (2x3090)
>>101787482gloopy
>>101786803Kek
>>101787490>>101787477Oh, makes sense.They managed to above and beyond to cripple their model.Amazing I guess.
>in the not so far future, anons will communicate with generated text, rather than type a post
>>101787517>They managed to above and beyond to cripple their model.its very easy for em these days.. they have the data set (probably took it with em from SAI) then ran advanced recognition AI over it .. every picture that showed genitalia was just removed, every picture that showed female nipples got a red blotch on the nipple (thats why you get nipples that look like gumballs) .. release save to go.. Russian idiots managed to make cunny necrophphilia in day 1 anyhow .. well lets see where it goes from here
>>101787523I'm already seeing reviews for stories I'm checking that use chatgpt.It's horrible shit.
>>101787557>Russian idiots managed to make cunny necrophphilia in day 1 anyhowWhat, for flux?
>>101787557>every picture that showed female nipples got a red blotch on the nipple (thats why you get nipples that look like gumballs)alternatively all they did was filter out nudity and the terrible nipples is what little understanding of nudity the text encoders havethe same applies to female celebrity names pulling some correct attributes, it's what CLIP knows of them
>>101787599go back in the archive to when Black Forest anon visited
>>101787557Yeah I guess, and they also got rid of most (female) celebs, characters, artists...The future of models seem to aim for more and more stuff deleted to please this or that pressure group.
I'm new to /ldg/ and running things locally, is this downloading a bunch of stuff or is it rendering an image? I'm doing it on my laptop so I know it's going to take a damn long time, but I don't even know if this means its downloading files or actually processing an image.
>>101787611no they surely have artistic nudity in the data set, but the data was censored in some way .. you can prompt nude photography artists you get their style and the poses.. does even understand stuff like "high key photography" (thigh gap) but the data that was allowed into set was sanitized, otherwise you can not explain the botched female nipples, it appeals to american financiers.. "no nipple was shown that is not male" kinda sad by German corp. but they need money I guess>>101787616nothing that can not be fixed by Autists with to much processing power
>>101787505
>>101787644you are downloading the basics.. lurk more, ask more some anons are friendly and help.. also what are the specs of your laptop?
>>101787644looks like downloading
Guys, I don't think she liked my presento :'(
>>101787652>you can prompt nude photography artists you get their style and the poseslike who>"high key photography" (thigh gap)what
>>101787652>but they need money I guessAt the same time you can get political stuff (trump), make it write nigger, etc.So it's not like the model needed to be sanitized.
>>101787670Miku doesn't talk like that.
>>101787665>>101787666Thanks anon. My laptop doesn't even have a GPU, so I know rendering this 256x256 image of cheese will take weeks. I'm doing it anyway.
>>101787687yea but for reasons unknown Trump and Tayolor Swift made into the dataset of Flux .. but Greta Thunberg didnt (SD15 knew her) pic related ancient SD15 gen of mine
>>101787714
>>101787720I'd gladly trade her ugly face for artists and nsfw.
>>101787714kek.. thats the spirit
>>101787698the fuck is she holding now? kek
What's the best guidance setting for text to show ?
>>101787731>>101787714Oh wow, I'm not sure if there's any point in waiting for this. I'm not up to date, but chances are Flux wasn't made to generate in 256x256, similarly to how SD1 sticked around 512, or SDXL with 1024.
>>101787778>What's the best guidance setting for text to show ?a high CFG seems to do the job better at rendering text than only guidance: https://files.catbox.moe/pggikh.png
>>101787798I see, I'll try that.Thanks
>>101787794>chances are Flux wasn't made to generate in 256x256Flux seems to be working fine un a lot of resolution ranges, even on high resolutions like 2048x2048, you do that on SD you get duplicated shit everywhere
>>101787827>even on high resolutions like 2048x2048How much vram does that eat?Man I wish someone worked on multigpu inference...
>>101787827Now that's interesting. Sounds like something about it's architecture must've made it more flexible then. I recall latents being saved in 256 for SDXL, so I wonder if they changed something about VAE decoding.
>>101787853>Man I wish someone worked on multigpu inference...There's that in a way, you can separate the models (clip, VAE, image model) into different GPU'shttps://reddit.com/r/StableDiffusion/comments/1el79h3/flux_can_be_run_on_a_multigpu_configuration/
>>101787279Top tier beauty
Just opened the oven door and found some hot and fresh...>>101787894>>101787894>>101787894
>>101787911
