Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101655488>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>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>Fluxhttps://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellhttps://comfyanonymous.github.io/ComfyUI_examples/flux>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
blessed thread of frenship
>clip art, spongebob squarepants drinking a vial of poison
official pixart bigma, lumina 2 and hunyuan finetune waiting room, now with flux 12b to keep us company
Repsoting the new kino model of the yearhttps://huggingface.co/black-forest-labs/FLUX.1-devWE ARE SO BACK BROS
just fucking load it in fp8--fp8_e5m2-text-enc --fp8_e5m2-unet
>>101671470LDG eating good
>the streets of new york with a bus badly photoshopped onto the road>LOCAL ANDYICHEb-bros?
>>101671493>just fucking load it in fp8>--fp8_e5m2-text-enc --fp8_e5m2-unet>unetWait that's not a DiT model? HOLY FUUUUUCK
>>101671493that command loads the image model + text encoder in 8bit right?
>hatsune miku inside the cockpit of a plane, she's looking at the viewer, smiling. outside the plane's windshield are the twin towers>>101671582kek
I used up all my gimmie on both sites It's over
>>101671601clear cookies then restart browser + modem
>>101670735Well, this model will not be for footfags, that is for sure. After testing it, it's only slightly better than SD3 for feet. Maybe something you can fix with finetuning, but for now you want to make sure your characters have shoes on.Other than that, pretty good. Remains to be seen how easy it is to finetune with these requirements. Also there is no training code or anything to get people started and it is not clear that there will ever be. Black box model with no documentation. Still infinitely better than SD3.
>woman lying on grass
>>101671236No flux gens?
https://huggingface.co/camenduru/FLUX.1->ae.sft>clip_l.safetensors>flux1-dev.sft>t5xxl_fp16.safetensors>t5xxl_fp8_e4m3fn.safetensorscan someone help a retard that will use Comfy for the first time of his life? do I have to download everything? what does those files mean?
>Flux can do intertwined fingersWE'RE SO BACK OMFUCKING GOD
>>101671535>not a DiT modelDOA
>>101671721If you are retard, then wait few days for retardproof youtube video.
>>101671732>can do hands>can't do feetIt's so over for footfags if those elements gets separated, because no one actually gives a shit about feet.
>>101671721read this, should help you https://comfyanonymous.github.io/ComfyUI_examples/flux/>A pale pink crescent moon glows softly, with a tiny white egg walking on its surface. The egg has sweet white cat ears and sparkling eyes, exploring the moon's gentle curve. Delicate, swirly patterns dance across the moon's surface, adding a touch of whimsy to the serene scene.
>>101671772P-perverts will save us.
https://replicate.com/black-forest-labs/flux-devthis is almost perfection
>a screen shot of a windows xp desktop with a wallpaper of an anime girl with big tits in a bikinibros look
>>101671850It's over for Stable Diffusion
>>101671850ah shit that's cool
Will there finally be support for multiple GPUs now?
>>101671772>>101671826if it can do hands, it can do feet, just some finetuning will do the trick, and desu it's not that bad at the moment, look >>101671841
this is so fucking cool
>>101671850>>101671903Can it do other operating systems?
>>101671850>>101671903Wow, that's actually pretty good.
>>101671850>>101671903How do SDXL and SD3 (lol) do with these kinds of prompts? I feel like the training data had to be extremely extensive and broad for it to pick up on stuff like this so well.
>trending on artstation, 19yo cosplayer of 2b from nier automata, short white hair, blindfold, black dress with cutouts and ornaments, standing on a piece of rubble in a ruined overgrown cityFrom the pro version
>>101671918>a screen shot of a linux gnome desktop, multiple windows with an image open combine together to make one anime girlit couldn't get what i was going for but that's probably due to my esl prompting
>>101671936>A view of a lake in the afternoon. In the side a part of a rusty tank with a canary with the beak open on top. In the lake a shadow of a large scaly creature emerging from the depthsGuess I'll be buying a 5090 next year
>fluxwhere the hell did this model come from
>>101671860SD was dead the moment sigma dropped>>101671977germany
>>101671893Dunno bro. Looks like feet are definitely this models Achilles heel (lol).
>a cute egyptian anime girl sitting on the floor reading a book, on the cover of the book it says "The holy quaran"
>>101671932this is a sdxl gen i made some time ago of a OS, flux is way better imho
>>101671959this is impressive, it's probably the best model at text, not even close
https://huggingface.co/black-forest-labs/FLUX.1-schnellwhat does "distilled" mean?
>>101672095>hatsune miku looking at the viewer, speaking. a speech bubble above her head says "yeah, it can do really long strings of words pretty well"
>>101672144yeah, seems like sky's the limit lol
Damn...
>>101672095https://github.com/city96/ComfyUI_NetDistbro, use this
>>101672150Footfags BTFO'd from image generation
>>101672220>6 months agoI'm not using that deprecated shit kek, but thanks a lot for it though, I never knew it was possible
>>101672220you can only use them separately tho right? i cant, for instance, load a model onto both and sort of retardely combine the power
ok got flux running locally on 4090, it indeed works with 24GB .. lets test that thing to its core, but wth, my fox has a different type of cake than the one in the example workflow .. is that cause fp8?
>>101672269no that's not how it works, those models work in layers, so one gpu will do the begining of the calculation, and the other gpu will end the calculation
>>101672290>ok got flux running locally on 4090, it indeed works with 24GBimage model + text encoder included? that's nice!
>>101672231see >>101672220
New flux model is insane. SD3 / auraflow and everyone else btfo. Hell dalle 3 btfo
>>101672311llava to flux
>>101672290what are the gen times? on my 4090 it's very slow
>>101672290you offload the text encoder in the cpu though right?
>>101672322I wouldn't say that just yet. Kind of feels like the model isn't as flexible/knowledge about styles as dalle.
>>101672329
>>101672344loras will go crazy
>>101672344It has extremely limited pop culture knowledge.
How do I add a negative prompt to that flow?
>>101672344it's not, but the comprehension seems way better. aesthetic has been a major issue in local since day 1.
show me foot fungus
>>101672379Closeups are easy. Do a proper "the pose". Google if you don't know.
Ok bros flux is looking really good. But it needs a way to finetune it. I've been pissed off for a while now at existing training scripts not being able to efficiently split a model across multiple GPUs (can't full finetune SDXL properly even with 4x4090). So I'm gonna make a pipeline parallel training script for diffusion models. I've already done this for LLMs, it's past time I do it for imagegen as well. Will start working on it this weekend. Note that I probably won't be able to get flux working end-to-end until they release training code / tech report, as I have no idea what the loss function is, details of conditioning, etc.
i'm using the flux website, how do i get the left image which is flux to look more realistic like the right image which is bing? both used the same prompt
>>101672400>Closeups are easy. Do a proper "the pose". Google if you don't know.how about that one >>101671841
>>101672433that's the fun part: you don't! welcome to local where the finetunes are always two weeks away!
>>101672425look at the comfy code
>>101672433Add 3D enforcing keywords. Mention lighting.
>>101672452Google "the pose". The problem is soles specifically. Closeups are easy, but very hard to get passable images of full body shots with soles in view.
>>101672400>this local ai is bad in this very specific pose, oh no, what shall we do?
>>101672472>why is my extremely specific fetish not in the dataset
>>101672454can't i'm on 1060 8gb vram>>101672457thanks that gave me a slightly better look
>>101672472>Closeups are easy, but very hard to get passable images of full body shots with soles in view.anon... the picture I just showed to you is a full body shot >>101671841
>>101672497interesting I never knew there'S 1060s with 8GB, always thoguht only had 6GB max
>>101672530it's integrated in a laptop so idk
>>101672456All the architecture is there, but I still think it is missing details of how to do training. Maybe all the pieces are there and I just need to learn more. My academic / math knowledge of diffusion is a bit lacking. Like what the fuck is "flow matching"? With SD1.5 and SDXL training is literally just mean squared error noise prediction, but I think with these models and diffusion transformers it's a bit more complicated.
How much longer until the shiny new toy loses it's luster?
>>101672474>>101672486The issue persists in every single pose I do. Feet are fucked, simple as. Requires finetuning.Also I am using the version you can use locally, not the pro version.
>>101672609your gens in general are fucked or do you not notice the overbaked quality?
>>101672594less than a week as people realize it's literally the same shit as kolors and hunyuan
>>101672625It's uncensored and doesn't require a Chinese translator, so no, it won.
feet are good bois
hmm>>101672625the more models we have the better
>>101672497you could try llava to describe your image and then modify the prompt further: https://huggingface.co/spaces/MBZUAI/LLaMA-3-V
flux web demo seems to be down or overcrowded. has anyone figured out how to get it running locally without it unloading the models every gen?
>>101672594It's going to be the new foundational model, the question is how soon until Lora and Finetuning support.
>>101672625Its actually better than sd3 api version so its the real deal this time. Also apache 2.0
>>101672662fp8
>>101672616I can tell. I think it's the model. I have no other knobs than my prompt and guidance. These are very basic prompts that work really well in all the other models.
>>101672716It's not the model, it's you twisting knobs because you're a retard with a foot fetish.
>>101672716>Scaphandra!
>>101672710how? teach me senpai
Not gonna install comfy trash for this, there has to be another way
>>101672725https://replicate.com/black-forest-labs/flux-devThere is only one knob to turn and I have not turned it. I keep guidance default 3.5.I am not using the pro, because it's not what you get to use. I am testing the shit you can use if you have 4090, which I don't.
>>101672741>>>/g/bst You know what to do
>>101672749There's this>>101671823Btw, I made a gif comparaison between fp16 and fp8
>>101672823fp16 looks like a ghoul
>>101672823>imgsli.com
>>101671489>requires rtx 3090/4090no we are not back, cooldown your hypefaggotry.
>>101672900no, but it can write nigger (which you are), so we are back.
>>101672900at 8 bit it can fit on a 16GB card. People might rig 6 bit which should not decrease quality by too much to run on 12GB
>>101672900oh I didn't realize that poorfags were the defining metric
why are imagefags relatively poor compared to textgods? with LLMs, 24gb minimum is expected, and it's normal to have people multi-GPU or create entire rigs out of 3090s. 70b+ models are normal. with image you make something over 3b and suddenly the entirety of mumbai is up in arms
Can any model be made to fit onto any type of GPU vram? or are there limitations to "quantization" or whatever it's called.For example in the future when there's an open source model that requires 32GB vram or even 64GB vram, etc, can they be made to run on a GTX 970 by scaling them back and losing quality?
>>101672994something you'll discover is many imagefags are underaged and aren't cerebral enough to read books
>>101672994the barrier to entry was simply lower
wheres meme anon he should be posting his b8 gens
>>101672625fucking retarded shill
jesus fucking christ this model has so much potentially architecturally but it's absolutely RUINED by the synthslop dataset. >Illustration of a gothic girlwhoever was responsible needs to be fired, it can only be considered sabotage at this point. so fucking close and the ball remains fumbled thanks to this shit.
>>101673056>a grey fish creature looking at a giant fish hook infront of him, the words "is this bait?" written infront of it, black background, monochrome
>>101672664>It's going to be the new foundational modedid the voices tell you this
>>101673099>>101672982
>>101673099Nothing a very extensive finetune can't fix. Gonna cost a leg and an arm to do however.
>>101672982No wonder SD3 sucked. Everyone competent left to make their own company.
>>101673112probably the whole "it's better than shitty SDXL" part
>>101673133so this explains why they were scraping midjourney or whatever? remember that shit, when the midjourney guy cried about stability scraping the site? they took all of that and dumped it into this model which is why it has that sloppa look to it. most unfortunate really
>>101673173its outright better than the api "good" version of sd3.
>>101673173pixart is better than sdxl and doesnt look like sloppanb4 muh text
>a grey fish creature looking at a golden fish hook infront of him, the words "this is bait of excellent quality" written infront of it, black background, monochrome
Come on China bros. Don't let some ex-stability rejects outperform you!
>>101673192Look I'm sure the Pixart Next version will be great and overtake Flux, but for the next few months Flux is what people will be using.
cool
>>101673204current pixart is better
pixartsexuals do not worry, we still have bigma up our sleeves
>>101673213Feel free to do apples to apples comparisons.
>>101673193
>>101673213lol, let me see a it make nude woman without using another model as a refiner
can we not offload the flux t5 model onto cpu like we can with pixart?
Guess I'm buying A6000.
>>101672982>https://x.com/EMostaque/status/1819037255974973867Last year there were so many articles about how Stability was going to die next year, go bankrupt, etc, and I thought "no way", but I didn't expect a dozen fantastic models popping up to eat their lunch. Of course they're going to die now, who gives a shit about Stability when we have all these options.But I swear the writers of those articles were just pushing anti-ai narrative. So those journos ended up getting BTFO too, since open source AI is in an even stronger position than before.
>>101672994>24gb minimum is expectedI've been using 8GB VRAM + rest in system RAM for like a year. Currently using a mix of Nemo 12B, Gemma 9B and some llama3 tunes (and some llama2 13B) and I'm pretty content since I just need fun stories instead of the model solving my python homework for me. Vramlets exist everywhere.Sucks that for SD Forge was the only good UI for 8 GB and the chinese cat abandoned the idea of it being a good version of Auto1111 and just made it into some gradio4 experiment. Nodes suck, fuck rearranging your entire setup for something that is literally one checkbox click in gradioslop.
>Painting of a unwashed indian man standing next to a toilet screaming at it. Painted in a classical style, detailed brushstrokes. Pierre Auguste Cot, Gustaf Wilhelm Palm, Franz Xaver Winterhalteryeah i'm not feeling the aesthetic. slopped up to the max, can't even doing a simple painting it seems. i tried with and without the artist tags, made no difference. i wouldnt crown this kind of local yet, stylistically it's just bad. still such a stylistic and detail difference between local and dall-e. what is the cause of this and will any team ever address it? midjourney somehow is still looking beautiful despite v5 being about a year old. i don't think it's a technical or architectural issue but rather the datasets. why are they so bad?
GOD the DETAILhttps://files.catbox.moe/ifr83d.pnghttps://files.catbox.moe/edl9ww.pngSo good.
>>101673315yeap, and there's still more models to look forward to. the local image gen drought is finally beginning to clear. while flux might be a bit slopped, it's still going to set a new standard, so i'm pretty excited to see what the hunyuan, pixart and other guys do in response.
>>101673356>flux can't into paintingsNo way kek
Don't mind me
>>101673377The girls are so skinny, I think I've been brainwashed by Pony's thickness.
>>101673315At the end of the day the reason why SD3 sucked is they post-processed lobotomized it. As Auraflow and Pixart has already proven, you essentially only need one or two people to train a model. I don't even think Flux is doing anything particularly novel and once you've decided your parameters and architecture, it's just shoving millions of captioned images in.
I'm only using the web version since I'm not made of VRAM but it feels to me like FLUX is good at certain things like text gen and making a good looking image overall but bad at other things like actually being able to fulfil my prompt. Many subjects, styles and mediums get ignored completely.
>>101673419>Painting of a girl, detailed brushstrokes this just looks like some greasy 1.5 shitmix really. what the fuck happened to the art? who is actually praising this besides jeets?
Some upload this to torrent
>>101673452>plastic barbie doll 3d modeleww
>>101673377WOW, can you give me the prompt anon?
>>101673356lmao same prompt
>impasto oil painting of a cat using a crystal ball to predict the future
we're so back
>>101673478barely passable painting, no where close to impasto
my friend next to me says someone should try to make something like this with flux
>>101673474just played with the default workflow one but:cute nude anime girl with long messy blonde hair and blue eyes sitting on a chair in a old dark victorian mansion with a bright window and very expensive stuff everywhere. Outside the window is a fantasy landscape with a airship in the far off sky
>>101673491its ogre, I asked for an abstract version and it just made a regular painting, >impasto oil painting of a cat using a crystal ball to predict the future, you can clearly see the heavy brush strokes and lumps of paint in the canvas adding a lot of texture
>>101673476I think Flux made a critical error of using AI captions for 100% of the dataset. I think having raw alt tags / search engine titles are a must to ensure a diversity of prompting keywords especially for teaching artists and pop culture.
>>101673478same prompt, i tried on base sdxl
>>101673478>>101673518>>101673173 your response?
>>101673491>>101673466>>101673356yeah I think it needs further training to add more variety on it, looks like flux-dev is worse than flux-pro in that regardhttps://blackforestlabs.ai/announcing-black-forest-labs/
>Impasto painting, cat portrait, thick textured brushstrokes, dramatic lighting, chiaroscuro, high contrast, detailed fur, expressive eyes, dynamic pose, oil painting style, palette knife texture, 8k resolution, in the style of Rembrandt and Van Gogh>>101673528idk, I just downloaded and am testing it, once I'm done I'll be back to pony and generate anime coomslop
>>101673504it might not be good at paintings, but it sure is good at minecraft
>>101673500https://files.catbox.moe/3wgvrw.png
>>101673515its prompt adherence is outstanding at least but yeah styles are bad. and it's not only ai captions, they had plenty of synthetic data.
>>101673528No one model is universally the best at everything. But for general memery Flux is top dog.
>>101673541I think anon meant that, with those examples, XL is superior
>>101673554cope
why are we having model wars?
>>101673569Flux being good isn't stopping me from training Pixart 1.3B. I'm sure people are going to criticize the shit out of it too for some use case.
>>101673561whoops, misread it>a wrirstwatch in the style of piet mondrian
pony's base aesthetic is also trash and needs loras to save it. Will be the same for this but with far far better prompt comprehension, colors and details
Why does ComfyUI unload and reload models everytime I make a gen? that's retarded
>>101673602you have got a skill issue
AHAHAHA you should've listened, fuckers. told you that synthetic data is a fucking mistake. now you're left with a model that produces nothing but generic ai-tinted garbage. and this will continue again and again until the faggots in charge do the sneedful and learn2scrape. shameful how 1.5 has more stylistic variety than the latest-and-greatest gigamodel. back to waiting for bigma (which will have the same issues)
>>101673578this, imagine being a fucking stan for an AI model. Bruh. just use whatever becomes the best
>>101673602make sure you are using fp8 t5, and also set the unet weight dtype to fp8
close your eyes everybody, i'm about to 1gir-ACK
>>101673635>make sure you are using fp8 t5, and also set the unet weight dtype to fp8that will stop the unload load bug?
>>101673628you the same dude shitposting on /h/?
Yea, flux will be THE base model. Rip SD3
>>101673648it's not a bug it's a lowvram optimization.
What I should download the FLUX.1-schnell or the dev?
>>101673648it isn't a bug, it will if you have 24 gb vram or whatever the amount is for holding the whole thing in VRAM, otherwise it will keep having to swap
>flux, the base model
>>101673669but that shouldn't be happening, I have enough ram and vram
>>101671236Know the fucking differenceFlux comes in two variants:* Timestep-distilled (`black-forest-labs/FLUX.1-schnell`)* Guidance-distilled (`black-forest-labs/FLUX.1-dev`)Both checkpoints have slightly difference usage which we detail below.### Timestep-distilled* `max_sequence_length` cannot be more than 256.* `guidance_scale` needs to be 0.* As this is a timestep-distilled model, it benefits from fewer sampling steps.### Guidance-distilled* The guidance-distilled variant takes about 50 sampling steps for good-quality generation.* It doesn't have any limitations around the `max_sequence_length`.
>>101673628tfw still waiting for anons 100% synthetic dataset model
>>101673690>>101673635Requesting a muscular man with THICC thighs please
>>101673706loras when, controlnet when, embeddings when
>migu
non distilled version is even better, who could have guessed. Gonna need to tensort this thing though to get useable speeds again.
>Classic book titled "The Most Dangerous Game", the cover showing a bunch of toddlers running away from a giant pitbull>FLUX.1 [pro] and [dev] surpass popular models like Midjourney v6.0, DALL·E 3 (HD) and SD3-Ultra in each of the following aspects: Visual Quality, Prompt Following, Size/Aspect Variability, Typography and Output Diversity.swing and a miss. better luck next time! the localgrift continues
>>101673711does this use a 16ch vae or something, it does the wires pretty well
>>101673706lmao
>>101673794oh man, the new local model that just came out today isn't doing as good as dalle for a specific prompt??? it's fucking over for local models forever...
>>101673790That's Pro?
>>101673756idk if really understands kek
impressive coherence but the style is extreme AI slop just like dalle3when are we going to get a model with that has the SOVL that dalle2 had. it had coherence problems but it could do art that felt like art
>>101673821https://huggingface.co/black-forest-labs/FLUX.1-dev
>>101673817kekd
>>101673833Flux unironically can produce images that are difficult to discern as AI. Dalle-3 has a weird rendering pattern that makes it obvious.
>>101673829damn, thanks for trying
>>101673833Yeah, I think its because they focus a fuck ton on aesthetic refining to get a good slop that pleases 90% of the people
>>101673844>Flux unironically can produce images that are difficult to discern as AI.And where did you see those? Please post some.
>>101673836But it says distillation.
>>101673851nta but it sure aint pleasing me. looks like generic revanimate trash.
Maaaadeeee the OP COLLAGE againnnnn
>>101673817i think i know how to use it now
>>101673833Jiust wait for LoRAs. Pony mixes with loras reached peak soul aesthetic already. This can be 10 times better. I'd rather have loras than built in styles.
>>101673860And thats what 90% of people like>>101673849
>>101673854And have you stubbornly say they're shit? Nah.
>>101673922I'll humbly accept your concession
>A buff muscular yellow Minion from Despicable Me wearing a black shirt that says "Never Goon"
>>101673949There's no concession, I can tell you're being a massive faggot right now.
>>101673954flux gen same prompt
>cctv footage of donald trump silently levitating in the sky, far far off to the distance, he is just a shadowy figure, it is a dark stormy night, grainy footage
>>101673954I am 1000000% expecting to see the minion on the right reposted at 20% quality with impact text within the next two weeks.
>>101673981>>101673954I'm retarded, pic here
>>101673954>zoomer cancer """memes"""Fuck off
>>101673985>cctv footage of donald trump silently levitating in the sky, far far off to the distance, he is just a shadowy figure, it is a dark stormy night, grainy footage, analog horror
>>101673954always goon
>>101674003gooner spotted
>>101673898>I'd rather have loras than built in styles.Kill yourself
>>101674014oh shit, I have this on vhs
>>101674014ahaha fuck that one is good
>>101674014I like it.
>>101673981
>>101673794Only domain in which flux is better than Dall-E 3 is portrait photography and maybe big scenic pictures of forests and waterfalls.Don't know why they have to lie and say it's better than DE3, when that is clearly not the case. If you however take into account DE3's lack of control and schizo filters, then anything is better.
>>101674020Underage faggot spotted.
Dear baker, please put some of these flux gens in the next OP The funny ones Please
>>101673802those are very nice wires desu
>>101674042post a prompt that you think did worse
>>101674042it has a cleaner look for photos and some stylesdalle looks like shit to be honest. but the style control is certainly better on dalle.
>>101674025Fuck off, retard. Built in styles are the reason why all dalle gens look like the same low effort slop.
>>101674075make her take a shit
>>101674075what? thats bad styles, not built in styles, even NAI had a lot of cool builtin styles
>>101674075>Built in styles are the reason why all dalle gens look like the same low effort slop.what a fucking braindead fucking take, rope loratardwasting time on baking loras for small time clout does not excuse your existencestop advocating for castrated models
>cctv footage of donald trump silently levitating in the sky, far far off to the distance, he is just a shadowy figure, it is a dark stormy night, grainy footage, analog horror, low resolution, the image is extremely zoomed inthis one was funny so i couldn't resist
>>101674067What is this weird flux fanboying in these threads? Why can't you criticize any aspect of it without having bunch of nerds attack you instantly.Angry employees perhaps?How about you do your best SFW work in flux, then pass me the prompt and I will Dall-E 3 it to BTFO you.New model is exciting and it's good, but don't be a mindless shill.
>A detailed ukiyo-e woodblock print illustration of Sora from Kingdom Hearts leaning against a maple tree, autumn colors. Traditional Japanese style.another style down... surely the synthetic data had at least a couple of images in this style? why the fuck is it failing at this?
>>101674042DE3 is pretty much outdated garbage at this point. It's been mogged by MJ in anything that is photorealistic since the start, and now prompt consistency and text is also better.
>>101674117Because automatic aesthetic filters will remove everything that isn't a high contrast high saturation image from the dataset kek
>>101674114do you work for microsoft / stability? Why so upset I asked for a prompt that dalle did better since your complaining.
>>101674112This makes me think about how many funny meme gens anon made with SD3 until we came to the conclusion that it's shit. That's a really funny prompt thodesu
>>101674104>he thinks lack of hardcoded styles means "castrated models".Lol, it means your styles are as shallow as a puddle, compared to absolute control you have with LoRAs
>>101674075absolutely freetarded take really. model having a vast arsenal of styles is far better than having to bake overfitted greasy loras. plus you can then combine loras with styles to enhance them even more. never understood why people wear these limitations as badges.
How I set f16 weight?
>>101674161>pleeaaaase rape models harder, I need my civitai downloads and (You)s!!!! without lora baking my life is worthless!
>>101674179because faggot wants his clout. a lot of faggots literally don't want to even gen, they just need a model to bake loras on. that's why they want this.
>>101674156>I asked for a prompt that dalle did better since your complaining.That would be to easy, since I could craft a complicated prompt which flux can't do and DE3 can. I am making it harder for myself. You craft the prompt and I DE3 it. Only slight modifications on my part.1024x1024, do your best or BTFO.
>>101674214I accept your concession
>>101674114It seems obvious you desperately want Flux to fail.
>>101674234I already accepted yours, since you can't provide your best
With Hatsune Miku, we will make america great again!
>>101674239No I don't. It's good stuff. I just want to be able to point out when it fails in some aspects. It's not perfect, but it's the best we got and probably will have for a long while.
>>101674179It's as useful as having random redditors rate output images when making SD model. Built in styles will never ever be as good as LoRAs made by dedicated people who can just steal any dataset and make it also uncensored. You are clearly a nogenner if you think that built in styles are useful.
>>101674268I've never seen someone so biased towards negative. It's pretty obvious. Flux costs you nothing.
>>101674259>ACK-sune miku
>>101674281>Flux costs you nothingCosts me 4090, huge amounts of electricity and my time. None which are free.
>>101674114>dalle plebs are this delusionalKek. also kek at "sfw prompt" part.
>>101674276This, I just hope no one trains any finetunes, I want to do a carefully curated lora for every concept, character and style. Ideally, we'd just use loras without the model.
>you vill enjoy de slop
>>101674295Go and use Dalle-3 faggot. I heard it still works.
>>101674276Fucking freetard, actually hang yourself. It's people like you that hold local back. Why don't you just bake a lora for every fucking thing then? Why aren't you still using SD 1.4? Who cares about comprehension why you can inpaint? Might as well just pick up a pencil too
>>101674295go cry about being poor somewhere else
>>101674308nothing's preventing you to continue the training of flux so that it has more variety anon, that's the beauty of local, you're not happy about something? you can fix it
cozy thread today
Go make me something like this with dalle
>>101674324He's so poor he relies on Bing Create's daily credits.
>>101674324>here is slop>fix it for me for free
>>101674114>flux fanboying>is the one shitting on new model for absolutely no reason????????????????????
>>101674333very cozy
>>101674349Are you the seething foot fetish anon?
>>101674349>here, we spent millions of dollars training that model, you can have it for free>NOOO IM NOT GONNA SPEND HUNDREDS OF DOLLARS TO END THE JOBunironically kill yourself
>Jeets come out of the woodwork to rabidly defend every flaw in their modelNormal people would go "yeah, this ls an issue. Hopefully it can be addressed for future releases. Does anyone know where to contact the devs?". Jeets go: "NOOO IS GOOD ALREADY it cant do style is good that means you are have to be more creative yes very goood". Actually crazy how they take every criticism personally as their pigeon-brains immediately start resorting to console wars instead of the actual mature approach of recognizing flaws and trying to figure out how to improve them.
>>101674314>Why don't you just bake a lora for every fucking thing thenThis is objectively the best way to do it, lazy faggot.Also yes, you are supposed to grab a "pencil" to make your controlnet fixes, you are supposed to know anatomy for artists and composition theory. This is art ,and you, lazy goyslop genners, are giving it a bad reputation.
>>101674368No. I am just shitting the thread for fun and fanning flames.
>>101674384>it was a bait all along
>>101674384this is why local is forever irrelevant. laughable post
>>101674377What's immature is the absolute fanboy seething whenever something comes out that threatens your toy.
Where is my dalle version at? Still waiting?https://files.catbox.moe/72lgal.png
>>101674377What you don't understand is that it's by far the best local model we ever had, of course we're gonna be greatful towards them. We never said it's perfect, and like I said, if you want to improve this shit, do the work, they have done easily 95% of the job, and you're bitching because you don't want to fill the 5% remaining, are you serious anon?
>>101674400>>101674403I accept your concession.
>>101674409flux nipples look very weird kek
>Painting of Mario in the style of Pablo Picasso wow it actually is really just that bad. it has literally zero concept of style this isn't even niche
>>101674414Hes either a troll or works for one of the other companies and is salty
>>101674423photorealistic nipples are okay, but it's really weakly trained on cartoon/anime nsfw
>>101674423I remember base sdxl's nipples looked very similar
>>101674409>Passes the multicolored ball testWe're so back
>>101674377Why don't you understand that you are not allowed to criticize this model!It's the best! Nothing will be better now or ever! If you ever say anything bad, you are Microsoft shill!
Imagine a model with zero synthetic sloppa thodesu
Why is this thread so anti-intellectual? You'd never find this kind of model worship on /lmg/. If anything they're more critical of their own models. Is this thread really just full of turd-worlder slop-worshippers?
>>101674439you could use her as a wheelbarrow
funfact:blackforestlab refers to the location of the company in the city of Freiburg, germany, which is located in an area called the black forest.
>>101674442Anon do you actually think you were being impartial? Holy shit even an 8 year old recognizes their bias. You came out swinging saying it's shit.
>>101674443SHUT THE FUCK UP AND BE GRATEFUL. They got you 99.99999% of the way there. If you want it done right, finish it yourself. Just shut up and generate you Microsoft Midjourney shill.
>>101674449All I hear is a dall-e worship, as if it was ever good.
>>101674414Whenever anyone points out aspects that any other model does better than flux, they get attacked.That is not normal behavior. We should be able to compare these models honestly.
>>101674377except it literally just came out, and probably only one motherfucker on this entire board is smart enough to actually contribute anything of value ON THE SAME DAY nigga please
>>101674429Kinda feels like there's a push behind the scenes against models that use synthetic data. Because it essentially clones them and destroys their business model.https://www.nature.com/articles/s41586-024-07566-y>AI models collapse when trained on recursively generated dataThis shit got so many comments and updoots on reddit.Where as this gets zero attention:https://the-decoder.com/ai-data-isnt-destroying-ai-models-after-all-researchers-say/
>>101674465not possible when you have a thread full of unironic stallmanist freetards like the mouth-breather above you. it's actually a mental disorder, they are incapable of recognizing flaws as long as it's free
>>101674465I won't attack you, I also agree with you that it lacks styles and varity, but that can be fixed with some finetuning, they made a base model that is so good it can be fixed by us, that's not the case of SD3 where NFSW and anatomy is completely broken, that's why we're happy of the result
>>101674465>anyoneYou mean you?You mean you, spamming the thread and seething because his obvious negative bias was pointed out?
>>101674480How is there a push behind the scenes when every recent local model uses synthetic data, and all how the same glaring issues? Holy shit now you're resorting to conspiracies to justify your unaesthetic slopbaking. Actually fucking pitiful how far local has fallen. Anything to try and skirt around adding actual art back into the dataset
>>101674492badass image
>>101674480tbf, reddit is very much anti-ai (in public) which is why anything that makes ai look bad, or "will be the death of ai" is gonna be upvoted hard. but also, reddit is a pretty bad measure because it's so astroturfed and botted
Can someone make a node with nevative prompts in it? I have no idea how to make it work
>>101674500how many times are you going to make posts like this? do you really think you can affect the trajectory of what developers are doing? it's getting pathetic
>>101674481>flawsNow compare the base sdxl with the latest pony finetunes, and tell me about the flaws. Who the fuck compares a just released open source models with censored corporate slop with zero adjustments.
>>101674518get make another clip text encode, get k sampler, replug everything into k sampler instead, then turn cfg to like 1-2
>>101674465which opensource model does what better than the opensoruce model flux? do you see what a completely degenerate piece of shit you are?they even called it dev so that every fool can understand what it is supposed to do.go to your garage, do the world a favor and hang yourself.
>>101674500All I see is someone with no skin in the game seething and schizoposting.
>>101674532pony is absolute garbage
>>101674423they do, finetunes when?
>>101674544Pony is the best model on the market.
>>101674481>they are incapable of recognizing flaws as long as it's freewe recognize flaws anon, but we always make comparaisons relative to other stuff, and compare this gem to the fucked up SD3 and you'll see how back we are
>>101674544b-b-b-b-but you can prompt penis inside a vagina with itwho cares that it looks like vomit
>>101674556Pony is a fluke and you'll never see a modern version of it.
I think everyone can now see that it is just a troll. Ignore.
>>101674537>then turn cfg to like 1-2why? flux seems to be working fine on cfg 8 here
>>101674480>https://the-decoder.com/ai-data-isnt-destroying-ai-models-after-all-researchers-say/>language model LLMs =/= imggen models, retard-kun
>>101674585distilled model means lower cfg needed. Depends if your using distilled model
Some flux friend could help me with this error in my ComfyUImodule 'torch' has no attribute 'float8_e4m3fn'
>>101674579this
schizo pattern
>>101674598no I'm using flux-dev, the better version?
>>101674625
>>101674635>>101674625i meant to say kek
>>101674589True, however that article is in fact responding to the other one I linked.And can you be certain that art training and text training would be that different?
good model but with serious drawbacks. not the great leap forward i was hoping for when i saw all the praise. incredibly beefy too, likely won't get any finetunes at all sadly. the lack of accessible finetuning hardware will hold back progress far more than the lack of base models, of which we have plenty. we've received a plethora of new models over the past year yet nothing meaningful ever evolves from them, they get ditched in under a month because the flaws are far too glaring and the cost to fix them is far too high. flux will join them on the wall of "man i wish this had a finetune!" within a few weeks as people go back to waiting for XYZ which will totally receive mass adoption this time.
>>101674664where does diffusion happen in imggen
>>101674656Don't do it anon...
>>101674664nta but because aesthetics matter for art, if a image model makes ugly gens it's pointless.
>>101674537Did I do something wrong?
>>101674671lora training should not be out of the range for most big tuners.
>>101674673>imggen*txtgen
>>101674671Oh boy, these flux fanboys are about to rip you a new one. You did something you can't do, your criticized their toy.
>>101674671Earth to anon, there will not be a small model you can use on your toaster again. Stop talking about Dalle-3, a model that runs from a datacenter.
>>101674683cfg 8 for one, needs 1-2 I found
>>101674683cfg too high
>>101674656What prompt? Is it a stereo effect?
>>101674664Synthetic text is only logically wrong, not syntactically wrong. AI images are still melted piles of garbage. >Syntheticettxt is only logical^^~8ly wrong, not not nnnnnnnnnnnnntically right right AI IMAGES melted sitll piles garbbagethat's what the text equivalent of the current ai synthetic garbage would look like. why people think training on fried 7-finger hands and nonsensical backgrounds is beyond me. jeets buying into the "ai powering ai future" hype i guess
>>101674656Interesting idea.
>>101674700>>101674703oh yeah it works now, why itwas working fine on cfg = 8 here?
>>101674681you see, you have what's called bias and delusion, you say it's ugly but that's objectively incorrect
>>101674725>but if i only choose the very pretty gens itll look good
>>101674677wdym?>>101674721yes
>>101674692yeah, seems the wave of complete tards emerged. guess i can at least laugh at them in 3 months when this model remains sitting at the bottom of a lake as they go back to begging for porn finetunes. don't want it to be like this, but seems i'm one of the few here capable of seeing the unfortunate reality.
>>101674730Ok ComfyUi noobs, I changed the nodes so that it has a negative prompt in it, you download the picture and you load it to get what you want
>>101674692>>101674743why did you reply to yourself
>>101674743Yeah the consistent theme I've noticed is loudmouth leeches bitch about people not doing things for them.
here, negative prompt workflowhttps://files.catbox.moe/2wckwm.png
>>101674753are you from /sdg/ and shitting this thread on purpose?
can someone explain to me why so many closed diffusion model worshippers are fagging around in our 'local diffusion general' instead of just creating their own 'closed diffusion general'? are they too retarded to read the three words?
>>101674757Doesn't work. Is it for castrated model only?
Just admit your lazy and don't want to source real images for your dataset
https://blackforestlabs.ai/up-next/ARE WE BACK YET?
>>101674813
>>101674835Just admit you have a negative bias and they need to make Schizoposting a 3-day ban
>>101674813freetard
>i acktually enjoy the slop
>>101674671>>101674753I think you don't have the hardware, and that is all.
>>101674884You are just making shit up in your diseased mind
>>101674467I should gen some Redwall with this.
>yes, in fact, deepfried ai images are actually good, just look at all these dalle clone checkpoints
>>101674671/lmg/gots are getting finetunes of 120B models, I think we will manage to finetune a meager 12B model
>>101674935
>>101674671I'm gonna try to make a training script that can train a lora on 2x3090 (>>101672425). And it seems like you can run it for inference with just 24GB, since the DiT can be loaded in fp8 with minimal quality loss. Open weights LLMs being much larger than this hasn't stopped the community from making a lot of finetunes. If the model is good and has potential it will happen.
>>101674935who are you talking to
It's been a while since we've had bread this fast but here we go...>>101674851>>101674851>>101674851
>muh hecking meme gens desu
>>101674935The fuck you talk about? it can make way better realistic pictures than dalle3
>>101674605For Anon with the same problem as me, remember to upload your tchor and trasformer in Linux to get this work.
