Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101720035>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
Blessed thread of frenship
>Blessed thread of frenship
So in a couple days we went from "Flux is impossible to make loras for and it will NEVER happen" to "it may be possible but you need 75gb of VRAM" to "okay just 35 but that's still nonviable" to 24 to 18 and now targeting 16 and maybe even less. It's just a matter of time until it's fine-tunable with a 3090. SD is over.
>>Blessed thread of frenship
finally, the miku series I have been waiting for.
Guys i'm scared... Momo drew this
>>101725162Okay but how is the AMD support
>>101725216She's asking for it.
>>101725162We still need to see a final product of one of those loras. There were a couple of cope settings to train a lora on XL with 8GB when it released, but there was clearly a quality cost attached.
What if daredevil was superman
>>101725236My experience has been that 12 gb is the bare minimum to train a good style lora for pony, so I frankly doubt that will be enough to train for flux
World of Warcraft ingame screenshot, UI and HUD visible, the main character is Donald Trump, he is wearing plate armor and is holding a large greatsword. The background is Icecrown Citadel from World of Warcraft. The action bar at the bottom has sword icons. The minimap at the top right is a map of Icecrown.
>>101725248he is
>>101725264
>>101725268>French fry burritoYou Americans will eat anything.Also Flux is good at food
>>101725216that's pretty good.. what's flux?
>>101725296i'm as American as you
>>101725331
>>101725331hey that's my mom
>>101725317>>101725341Love the naga
>>101725286Miku edition:
Upscale from 1024x1024 to 2048x2048 rather than directly genning at 2048 is fine. But fuck me is it slow.
>>101725396We can upscale with flux huh? Can it inpaint too?
>>101725396Upscale using?
>>101725396Forgot pic>>101725403Haven't tried inpainting mostly because I hate inpainting in comfy.>>1017254094x ultra sharp
>>101725422>>101725396>Forgot picI am a retard
>>101725433did you denoise it after upscaling? if so by how much? how much more vram on the upscale?
Swimsuit made out of guts
>>101725433Yeah ran another ksampler at 0.2 noise with the same seed
>>101725030Been running into all sorts of issues while trying to get sd working locally on my machine (AMD/Linux). The guide on the stable-diffusion-webui-amdgpu repo linked to in the rentry up above has gotten me the closest so far, and while the webui does now install/launch without errors, I can't seem to get it to actually start generating anything. The buttons in the webui seem to not work and there's no new output in the console after start up. Any ideas what might be causing this?inb4 no model, this guide installed v1-5-pruned-emaonly.safetensors for me, but I also tried two other models from civitai without any luck.
>>101725473see >>101725516I am fucking brain damaged today I don't know what is wrong with me
>>101725531You could try comfy if auto is flat out not working. I don't use AMD/Linux so I have no clue what the issue might be. You should ask /sdg/ too in case someone there does.
>>101725531Are the other models you are trying also 1.5 models or are they something else?
>>101725399smother me in them titties
>>101725631One of them is based on SDXL 1.0, but the other was based on 1.5.
>>101725531You have to get amd working right first. Then you have to get Python working right with Pytorch.In Ubuntu, you have to use a venv, or at least I do.
I'm doing a 2-step high res images, but the second sampler adds this shitty looking noise, what is fucking up? Both samplers are euler, normal, 4 steps.
>>101725701The guide I followed created a venv and I am makings sure to activate it prior to launching webui.
what a time to be alive, and they said dall-e would never be open source.
>>101725699get the 1.5 working first, I did.
>>101725714did you install everything with pip from install.txt like it said?
what guidance do you guys use for flux? I'm still not clear on what changes if you raise or lower from the default 3.5
>>101725702Was your guidance extra low when you created the first image?
>>101725719pixar style:
>>101725739It was at 1 yes
>>101725743
>>1017257342-2.5
>>101725727Is the v1-5-pruned-emaonly that came with my install not 1.5? Regardless, I'll give manually downloading it a try.>>101725731The install script (I believe) should handle that for me, but regardless, I already tried manually installing all the packages using the install.txt, but I just get the same result.
We modeled our armor to cover the areas where soldiers returning from battle are most commonly wounded.
hmm
>>101725765neat
>>101725781How do you think they made it safe, Anon?Also, they are definitely still in there. Only their names were scrubbed.
>>101725781>>101725800
>>101725781>suit, vans/converse shoes
>>101725822
open source wins cause dall-e would reject this prompt.
>>101725762I downloaded exactly the v1-5 whatever and put it under ComfyUI/models/checkpoints
>>101725800names scrubbed how? you can prompt for swift
>>101725860>>101725762
>>101725872Not everything was scrubbed.Just like you can get nudes yet they scrubbed most of them.
>>101725877Can you reverse it? A giant cockroach with a plate of humans?
>>101725872You can probably gen taytay because her name is ambiguous, not clearly female. I have a feeling their VLM just scrubbed female names, rather than all celebrity names, so it's harder for us to gen them fucking donkeys and harder for BFL to get sued. Nobody cares if we gen a horse cock onto George Clooney.
it knows what trollface is but they pruned pepe the frog from the dataset. sadness.
>>101725895do you prompt pixel art at the start?
>>101725895Neat, prooompt?
>>101725531I was having the exact same issues (on debian), switching to comfyui solved most of it and from there you can just steal someone's workflow and build off it yourself.Had to install pyenv and python 3.10.6 through it first, then set pyenv shell 3.10.6 before installing anything else related to comfyui. automatic1111 wouldn't even recognize my gpu no matter what I tried.
pyenv shell 3.10.6
>>101725916>>101725933https://files.catbox.moe/kcyt48.pngThe prompt is too long to post. I accidentally pasted the prompt I gave the text gen AI as the first line.
>>101725888https://files.catbox.moe/kud2ta.pngLooks more like barbie dolls and cockroach wasn't that huge, so it sort of works. Maybe I need to try it in a different context.
>>101725995Pretty funny though, maybe realistic cockroach man?
>>101725755how do I get stuff like this?
>>101725888nta and not really humans
>>101726008give it a style like pixar and describe the text and locationprompt is like:The title "POO in the LOO" is visible in bold and playful text. The background is Mumbai, India. Several Indian men wearing a Turban are squatting in the street. Poop is scattered around the road. The poop looks cartoonish. The image style is like a Pixar film. At the bottom the text "just loo it" is visible. Include Pixar branding to make it look like a Pixar movie poster. The Pixar logo is visible at the bottom of the image.
I love flux
>>101725531I haven't tried webui but I have no issues with comfy. What AMD card are you using?
the custom children's style fonts are making me laugh, lmaoThe title "Dora the Auschwitz Explorer" is visible in bold and playful text. The background is Auschwitz. Dora the Explorer is standing in front of Auschwitz with a smile on her face. She is running towards the entrance. A sign saying "Arbeit Macht Frei" is visible on a steel gate. The image style is like a Pixar film. The Disney logo is at the top of the image. Include a tagline that reads: "let's explore!" prominently on the poster. The visual style is in the style of a Disney movie.if it can do edgy prompts, anything is possible.
>>101725888>>101726025improved it
>>101726074do you prompt pixel art at the start to get that style?
>>101726084Amazing.
What the fuck will this look like in another couple of years
>>101726098b& for your own safety. You'll be able to rent AI time from trusted companies.
>>101726030for some reason AI doesn't like the word pedophile
coming home after a hard day of work to a plate of steaming humans
>>101726124lmao top kek
>>101726096yeah i just began the prompt with pixel artit's a gamble as to what magnification you will get, since these models are not trained to be pixel perfect
>>101726140Nice, the size is right!
sucks when it's almost right and you gotta wait a while to gen another try
I might cause an international incident if this is posted
>>101726166lmao it will do anything.
>>101726162
>>101726162what GPU u using m8? how long your gens take?what do you do while it generates?PC is basically unusable for me when it generates.
>>101726173yep, and this is why open source always wins. even the new llama 3.1 LLM models have uncensored versions now.what a time to be alive, huh?
>>101726063>if it can do edgy prompts, anything is possible.It can't do nsfw.
>>101726189rtx 3060, 2-3 mins for a 1024x1024 gen, PC is just fine while genning. If your PC is freezing up I suspect you are running low on RAM, which is required by the large CLIP models for flux.
>>101726227it can't do full frontal or porn. It does erotica and gore just fine
>>101726227Neither could any of the other base models beyond ass and titties.
>>101726236I use a 4090 and have 64GB of RAM, PC isnt actually freezing up but browsing with 100 tabs in chrome is somewhat choppy and cant really play any videogames or watch 4 videos on jewtube while waiting.its a pain in the ass.
>>101726227i don't see the appeal of nsfw that much, the images don't turn me on because i know they're not real
>>101726274I think its alright that it cant into nsfw for now, it forces people to be more creative.
>>101726227if you really want tits, just take a flux gen and inpaint/img2img with sdxl or ponyxl models.
>>101726271that's when you bust out the laptop for your second screen
I remember struggling with bats on SD and eventually putting them into negatives, flux gens them very nice even as pixel art.
>>101726236>>101726271these are my settings and I get good gens in 25-30 seconds with a 4080 and 32gb ram. fp16 weight kills efficiency though and you definitely need a 4090 or more than 16GB VRAM.
>>101726293for sure, I gotta upgrade my second PC.
open source always wins.>noooooooo you cant prompt that!>please pay for tokens to generate images!
>>101726311even 24GB vram isnt enough nowadays.
>>101726311that's pretty good, i measured 24 seconds per image on my 3090 with fp8
can't get it to do water, wetness, moistness, humidness, or anything like thatno ESL bully please
>>101726311what about all the other comfy UI settings like steps, samplers and schedulers and shit?any good combos?
>>101726311I don't get it, everyone keeps saying you need lots of VRAM but I never have issues genning full fp16 flux1-dev on my 3060. Kinda slow at 2 or 3 mins per gen but I can hardly complain considering the quality.
I fucking love flux. I don't even know if finetunes are absolutely necessary as long as we get LoRAs and Controlnets.
>>101726311i found changing the dtype from default to fp8 ran like absolute shit on my 3080. from about 1min to 10-15min. ram usage went way down though.
KINO
>>101726384I didn't mess with anything else. 20 steps for dev model. If you use the schnell model, you can get results in 1-4 steps even, think of it like SDXL turbo.
>>101726393Are you seeing >Loading in lowvram mode popping up?
>>101726394There are some loras on HF. So I think its doable. I'm just wondering if we can get a quantized versions. Q4 Q5.
>pulls you over>"booboo gaga gugu"
>>101726403Schnell only saves me about 10~20seconds and the quality (especially in text) is far worse for me. Unless I'm doing it wrong and need to use a different sampler or something.
>>101726425I think the dev model but with fp8 weight is the best mix of speed/quality so far, imo: but I don't have 24gb VRAM.
>>101726406>Loading in lowvram modeI see this and I run a 4090, apparently 24GB is "lowvram"
>>101726451We all get to be poorfags together.
>>101726227It can, it's definitely in there.
>>101726443I'll be perfectly honest. The FP8 outputs aren't that much worse than the FP16 ones. It's like a line here or there difference.
>>101726406Yeah, but that comes up for fp8 as well and the speed is basically the same as fp16. And terminal output confirms it is running in bfloat16
>>101726173yes, anything indeed
>>101726303did yu prompt thes hirt
this is what DALL-E doesnt want you to know:
>>101726484ye, just put "wearing a pink floyd shirt"
>>101726472Just remnants of it, at most artistic nudity or missed images from the training data set. And besides, nipples are fucked, there is almost no nude genitals, almost no erotica (as in lingerie etc outside of tame magazines).
>>101726406Not him, but I do (3060), also my CUDA graph looks like this
>>101726425>>101726443Same experience here on my 2060 12GB. 460 seconds on Schnell vs 470 seconds on dev-fp8, much more accurate on the latter.
>>101726530like castlevania on the NES, cool
>>101726471really shows how much technology is lagging behind for all this AI shit.
I love that you can just prompt for action and poses with Flux, no control net. No more 1girl staring vacantly into the distance>>101726511It's fixed with fine tuning. Base 1.5 and base XL suck at nudes too. You can wait for the finetunes or you can continue to bitch like a little faggot.
you are literally only limited by your imagination.
>>101726556needs a punchier title, the line itself can be the tagline
>>101726556>limited by your imaginationironic
>>101726555>It's fixed with fine tuning. Base 1.5 and base XL suck at nudes too.We'll see if people even can pay for that.>You can wait for the finetunes or you can continue to bitch like a little faggot.I'm just sad that it could have been amazing from the get go anon.I tried asking for characters to have fun poses, it almost recognizes no one, artists it's the same, even public domain ones.
>>101726591If you can't prompt without naming artists you are a promptlet, simple as that.
>>101726556>Pixar style>Edgy Alex Jones or jews or nazis or nigger textYeah.
>>101726590if it can generate stuff that dall-e rejects cause of "ethics" then you can literally do whatever you like, even generate isometric videogame graphics.
>>101726598Not what I wrote.
im impressed by the reflections desu
>>101726556Imagination is overrated, just ask mistral nemo something like>very detailed description of a very insensitive, offensive, crude and politically incorrect videogame cover, specify the title, the slogan and their placement
>>101726630Very nice
>>101726591I dunno what to tell you man. It is a BASE model.We have this same argument literally every time a BASE model comes out. And every time, every single complaint is fixed with tuning. At what point do you give up and realise your complaint is both a non-issue and stupid?
>>101726639
>>101726663This gives a planet of the ape timeline vibe, good stuff.
what a lucky man
>>101726687this is where white people came from
>>101726694I believe it
>in game screenshot of fallout 3, the main character is Miku Hatsune
>>101726675This looks fantastic, the fact that there's coherence between those little spike cluster sprites is amazing. Catbox?
It's him, the 'ecker
pixel art. Final Fantasy 6 in game screenshot. UI and HUD visible. The main character is Hatsune Miku. The setting is a medieval castle.
Will you be joining the competition?
>>101726668
>>101726714>>101726707Can you generate something that isn't symmetrical
pixel art. Chrono Trigger in game screenshot. UI and HUD visible. The main character is Hatsune Miku. The setting is a forest.
>>101726514bottleneck because leaking to ram
>>101726704https://files.catbox.moe/6hv0a0.png>>101726740It all depends on whatever wall of text I get from mistral nemo.
>>101726209Nice, which version is this?
>>101726785llama 3.1 uncensored, download LM studio (app to load the models, it's small) and then find it in the file browser, or google llama 3.1 uncensored models.
>>101726773Thank you bro
>>101726778hogs of war reboot?
>>101726804
>>101726483
>>101726796*Also you need 50gb of vram to run the good models at a reasonable speed* teehee
>>101726822nah you can get the 8b model that is like 4 gigs and it works great, dont get the giga hueg ones.
>>101726832Nah the 8b ones just don't get it. Sure they can do basic tasks but they can't understand the situation in the same way the larger models do.
>>101726548Cool, what proompt?
>>101726282creative, exactly my humor.there's a reason why 95% of the SD ecosystem consists of porn.now people are generating thousands and thousands of memes for 2 weeks that nobody cares about because everyone thinks their meme is better. *slapno future without porn
>>101726556Did you like The School That Never Was? Well, Alex Jones is back with his sequel, Frogwater.
I'm trying to generate a Ted Kazynski funko pop, but every time it spells Kaczynski correctly the image is incredibly low quality. I probably should have added "high quality" to the prompt though
>>101726865I'm not against generating porn tho>image limit reached
:( I can't use Flux anymore today, its limiting me now for 8 mins. I only have AMD Ryzen 7 5700G 57 °C Cezanne 7nm TechnologyRAM 48.0GB Dual-Channel DDR4 @ 1196MHz (16-16-16-39)Motherboard Gigabyte Technology Co., Ltd. AX370M-DS3H-CF (AM4) 41 °CGraphics 12272MB ATI AMD Radeon RX 6750 XT (Unknown)
>we hit image limit for onceBAKER!
>>101726856>A page taken from the Necronomicon written by Abdul Alhazred. The page is written in arabic along with various occult sigils and inscriptions in an unknown occult language. There are diagrams providing ritual instructions for communing with the Great Old Ones. The page appears authentic and ancient, and evokes a sense of indescribable dread.
Baker baker baker man, bake me a thread as fast as you can.I will not bake it not by myself,someone else bake it and bake it top shelf.
Baking.
>>101726952Thanks, Marmandy Jones.
>>101726834lmao I love this
>>101726952kys
fuck that guy, I'm actually baking
>>101726630engaged in Gorilla warfare
yo niggas who bakin?
>>101727187120s
I'm too baked for this
Not baking
Is this how /ldg/ dies?
These guys are trolling. I'm making the collage and will bake as soon as I finish.
>>101727285ty!
Bump
I'm baking, hold on!
bake is up
Need more bake-on
Baker?
https://www.youtube.com/watch?v=AeUwikDrA9I
>>101727444>>101727444>>101727444Emergency bake