Discussion of free and open source text-to-image models Previous /ldg/ bread : >>101965917 >Beginner UI EasyDiffusion: easydiffusion.github.io Fooocus: github.com/lllyasviel/fooocus Metastable: metastable.studio >Advanced UI Automatic1111: github.com/automatic1111/stable-diffusion-webui ComfyUI: github.com/comfyanonymous/ComfyUI Forge: github.com/lllyasviel/stable-diffusion-webui-forge InvokeAI: github.com/invoke-ai/InvokeAI SD.Next: github.com/vladmandic/automatic SwarmUI: github.com/mcmonkeyprojects/SwarmUI >Use a VAE if your images look washed out rentry.org/sdvae >Model Ranking imgsys.org/rankings >Models, LoRAs & training civitai.com huggingface.co aitracker.art github.com/Nerogar/OneTrainer github.com/derrian-distro/LoRA_Easy_Training_Scripts >Flux huggingface.co/spaces/black-forest-labs/FLUX.1-schnell comfyanonymous.github.io/ComfyUI_examples/flux >Pixart Sigma & Hunyuan DIT huggingface.co/spaces/PixArt-alpha/PixArt-Sigma huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiT huggingface.co/comfyanonymous/hunyuan_dit_comfyui Nodes: github.com/city96/ComfyUI_ExtraModels >Index of guides and other tools rentry.org/sdg-link rentry.org/rentrysd >GPU performance vladmandic.github.io/sd-extension-system-info/pages/benchmark.html docs.getgrist.com/3mjouqRSdkBY/sdperformance >Try online without registration txt2img: www.mage.space img2img: huggingface.co/spaces/huggingface/diffuse-the-rest sd3: huggingface.co/spaces/stabilityai/stable-diffusion-3-medium >Maintain thread qualityrentry.org/debo>Related boards >>>/h/hdg >>>/g/sdg>>>/e/edg >>>/d/ddg >>>/b/degen >>>/vt/vtai >>>/aco/sdg >>>/trash/sdg
>>101967881What even is this mutant model? A merge of Schnell and regular Flux? Who would want this
no collage edition
>mfw
There were so many good gens in the last thread too :/
lol suffer schizo
>>101967945>AbdallahAlswa>Adel_AII wonder where are they from?
>>101967956uhhh sir...?
So i have 6GB VRAM, 16GB CPU RAM on a laptop, how do I run FLUX on forge?
What is the best way to batch interrogate WD tags for ComfyUI? I still uses A1111, but the extension is outdated and doesn't have the new models.
uhmm, i didnt prompt that thing in the back, who's that wtf
>>101967999thats the latent fren. he hangs out in latent space and sometimes hangs out with your 1girls.
>>101967932is AI bad for humanity?whn will AI art replace furry porn artists?
Around 30secs per Flux gen on a 4090.Can I do better?
>>101968019Can you fuck off back to your containment thread schizo?
>>101967977Get isekaied into a family with more money.
Miku Hatsune, 6 panel comicdidnt even specify what happens in the panelshttps://civitai.com/models/656307
>>101968048Don't reply to him just mock him. He's obviously in a dark space. Also use the filter in the rentry
>>101968021bad AI is bad for humanitygood AI is good for humanitythe major problem isn't AI, it's that the people developing AI don't take alignment seriously. they're all rushing to be first to market that they're going to make a paperclip maximizer that melts us all into raw materials for paperclipsso what I'm saying is: invest in paperclips. huge gains
>>101967999She's needs my help
>>101968076Go back to lesswrong
>>101968034you could masturbate while looking at your 4090 while it generates.>>101967999inpaint that shit out. flux really good at inpainting. also inpaint her chin away, in a seperate take, with a non flux checkpoint.
>>101967977use a dev-schnell mergeit only needs 8 steps for a good gen
Flux has a poor level of artistic detail and I got proof, on picrel right is Flux generating at 256x256, on the right it's a 1024x1024 one downscaled to 256x256.Notice how the right one is much better? If the detail was good then the left one would look better! Just like in SD1.5 downscaling a 512x512 to 256x256 would make it x4 more detailed.It if was good you could zoom in to any 256x256 square on the original (>>101966997) and see level of detail like in the right picture, instead, the detail DECREASES!Both are Flux which means it's capable of much higher levels of detail but it's not giving it.To attain it you would need to split the 1024x1024 image in 16 256x256 square chunks, and use img2img on them to get the detail, and then stich them together, which sounds ridiculous.
>>101968094Anon... I... Check out tiled diffusion. It's been like 2 years since people automated something like that.
>>101968021I'm a furry porn artist and I have more customers than ever.A lot of my customers already make themselves content by the thousands of pictures using AI and they're good at it.But they don't get any kind of a dopamine rush from hitting generate, therefore my business will never end. Also my ideas are superior to theirs.Worst that can happen is, that I'll be the guy using AI to crank out more content for them than ever before.
>>101968094>on picrel right is Flux generating at 256x256, on the right it's a 1024x1024 one downscaled to 256x256.wait, which one is left and right again?
>>101968142>I'm a furry porn artist
>>101968142>But they don't get any kind of a dopamine rush from hitting generatewhat psychopath doesnt get a hit i almost dont believe it >Also my ideas are superior to theirs. based fuck normies
>>101968090Okay, so,where do i find that merge?Do i simply put the merge on the models folder?Do I need to download a vae and txt clip?
americans going to bed, neelesh spamming his overbaked biatch. shrug.>>101968132just ignore this fucking idiot. the earlier he leaves for good, the better.>>101968142based.
>>101967945Schizo ai clout chasers
>not a single bigma mention the past 4 threadsis it dead? did flux kill the griftbakes?
>>101968094>To attain it you would need to split the 1024x1024 image in 16 256x256 square chunks, and use img2img on them to get the detail, and then stich them together, which sounds ridiculous.how do we tell him
>>101968054
>>101967945Merges are becoming popular and all of them are claiming to have reached dev quality with schnell. Just a vramlet cope and clickbait:https://civitai.com/models/618792/nepotism-fux-devschnell?modelVersionId=698731
>>101968168>where do i find that merge?https://civitai.com/models/630820/flux-fusion-ds-bnb-nf4-gguf-q4-q5-q8-fp4-fp8-fp16-4-8-steps-aio-and-unetI use Q4 ggufyou can search "flux merge" for other merge>Do I need to download a vae and txt clip?yes, put them in the vae and clip folder
>>101968191nvidia bought it so they could turn it into a corpo product, probably
>>101968051Can you use azure?
>update comfy >use image metadata from july >output looks different Is it just me?
>>101968312same seed?
>>101968312had some weird deviation shit happen with certain setups in the past, maybe AYS related. should't happen tho. changed your hardware?
>>101968349>maybe AYS relatedayyy lmao?
>>101968054Words look realistic though for those who can read Japanese it's probably just gibberish
>>101968318Same everything, yes. The only weird thing I did was switch out that dynamicthesholding file, but this workflow doesn't use that. I thought it was something with flux but then I tried a July workflow which is XL and same thing happened. >>101968349Same rig, same everything (I think). All I did was update comfy.
>>101968352"it was aliens" scheduler.https://www.youtube.com/watch?v=TnWHIHveRxU(don't watch this and make the mistake of checking out various footage to see if that guy is a lunatic)>>101968170wouldn't mind another one anon.
<lora:water_flux_lora_v1:1> Miku Hatsune, 6 panel comic. in each panel, Miku is singing at a music concert with a microphone.>>101968360you can actually specify what happens in each panel and it will work, it's neat.
>>101968364>dungeon entrance>dungeon entrance, zoomer
>>101968397bottom left is just adorable. and prompt comprehension is "it was aliens" tier indeed.
>Just one more prompt and I'll go to bed>Fuck
>>101968377
>>101968457>fill the queue>go to bed>wake up to new gens>have a wank because it was all late night coomer brain prompts
>>101968457funny. I never do that. I either do single takes or set it to "keep going".>>101968482you're the best. even has it all. burgers, pantsu, you name it.
>update comfy>gens go twice as fastsometimes good stuff happens
>>101968633I'm going to shit myself in anger because my outputs are now even more sloppa.
<lora:water_flux_lora_v1:1> Miku Hatsune, 6 panel comic. in each panel Miku is wearing different stylish outfits.cute!
>>101968633>update comfy>2 essential nodes in my main workflow break, effectively bricking thousands of gens.sometimes bad stuff happens
>>101968507>has it allDoes it?
>>101968678I try not to lust over other ppls gens too much but you are pushing it. the thing is, I am hungry like a wolf and its 5am and I'd murder for a burger.
wer
>>101968708Do you yearn?
flux inspired me to try training a lora for the first time, came out reasonably well and took about 3 hours on a 4080. I'm using the kohya cli and wanted to see if anyone has any tips for success. Basically just using the suggested launch arguments in the github repo, with the 12gb vram optimizations, but as far as any other arguments not sure where to start. Did 1 epoch w/ 1600 steps, just copying another anons stuff from yesterday, but wondering if there's anything better to try.Used joycaption to bulk caption my dataset, not sure if that was ideal and it prevented me from having any real trigger word.didn't crop anything, just like kohya bucket them. Set resolution to 1024,1024 for shits and giggles
>>101968723What's with the pajeet?
>>101968739What's with the racism?
bro 256px images
>>101968746Don't be such a faggot.
>>101968699the texture of the pants is lovely. I want to touch them
>>101968723Train at 512x512 increase steps
>>101968708https://files.catbox.moe/23pd0g.pnghttps://files.catbox.moe/7ypt09.png
Q4_K_S runs surprisingly fast on my shitty system with a 3060 12gb and 16gb ram but if I run Q5_1, which is a slightly larger model, every time I change the prompt, it hangs for about a minute on the sampler node before it starts genning, what's happening there? It seems to be unloading something and then reloading something else
>>101967946>anon posts image>me: what if I...
>>101968678>>101968719yes. that is exactly the right word.
used blue archive lora, got neat scenery for first gen
>>101967946divinity original sin vibes
SAAAAAAAAAAAAARS
>now prompt migu
>>101968849What's with the pajeets?
>>101968888can you stop being a fucking racist
>>101968849>Still have butt-chinsWill not be redeeming this lora saar
>>101968771appreciated. just wondering how many gens of this particular toon you made. like, total, ever.>>101968887bring it.
>>101968888make them white if it bothers you so much. im not gonna do it for yoiu
>>101968898Why are you trying to stop me from expressing myself? Is that okay? To just prevent people from expressing themselves?
>>101968849this donation thing is gonna be the end of free models on civitai, they're obviously nudging you towards paying for shit
>>101968888kek
for those who decided to do a "update_comfyui_and_python_dependencies.bat" update and have slow gens, never do this, it update pytorch to 2.4.0 and makes everything slow and bugged, if you have done that I'd suggest you to downgrade to the original state which is 2.3.1, I went from 4.5s/it to 3s/it with that downgradehttps://github.com/comfyanonymous/ComfyUI/issues/4447
>>101968913I should try cardosXL, I really like your gen mate
Pony sloppers OUT
>>101968849Words cannot describe how little I desire to have anything to do with this lora... let alone pay for it lmao
>>101968917they can all go suck a large schlong.>>101968933we've reached the point. "indistinguishable". oops.>>101968964old gens, old upscales, sorry for mediocre quality & ty
>>101968993Mind sharing a catbox for one of them? They are superb.
>>101968960This is a windows specific issue
>>101968781If you mean the initial load it's because of mmap, so the "instant" load time you see at the node is actually fake and it takes a while to load it from disk directly to the GPU.(If you mean the prompt change thing it's most likely unloading the model to load CLIP/T5 onto the GPU to encode the prompt before loading the model back.)
>>101969013So that issue is relevant to 70% of users yes
>>101969017no it's not the initial load, it's for every single change made to the prompt AFTER the initial gen.
>>101968898POO
>>101968913How many gens?I don't know. her first true gen was 16,786.
>>101968898>>101968914Sirs
>>101969003thats a forge gen, here: https://files.catbox.moe/k3k87o.png>>101968960some friendly anon helped me upgrade to pytorch nightly (can also downgrade if you like), just check out one of the earlier threads, the info is close to the start of the thread.
>>101969053
>>101969059?
>>101969023what happened in nov.23?
>>101969044kek
>>101969028Could pin CLIP to the CPU or a second GPU if you have one, that's what I've been doing. Sure it's slower but not as slow as unloading most of the time. Just don't pin it to cuda:0 that'll OOM.Custom node at the top of this page for that - just add it after the CLIP loader https://gist.github.com/city96
Hello sirs, I thought the joy caption interface was lacking a bit so I tard wrangled JeetPT into making some changes.I gave it the ability to batch caption images from a filepath, you can edit the prompt for the llm in the window and there is a second tab where you can go back and edit the prompt for each image and quickly insert a bunch of trigger words of your choice if you want them.Interested?
>>101969075Yes, very much
>>101969069Stay in your containment thread please.
Have the flux team spoken at all since release? Do we have any info on how the model was trained/captioned?
Literally me genning
>>101969072thanks
>>101968938^ America 1990s< America 2024
>>101969055ty I love you
>>101969086https://pastebin.com/p2PWgJZKThat should work. Just past that script over your old app.py and replace the "FILEPATH TO YOUR MODEL" section with the folder you keep your llama 3.1 model in. I might add a gallery to the editing section later for ease of use.
These threads are moving so fast nobody will find out I have an average size penis.
>>101969100Nope, only bit of communication with them was their announcement about powering the image generation in Grok 2. That's it. Which is great, we don't need all the community management BS Emad was doing.
>>101969100>Have the flux team spoken at all since release? Do we have any info on how the model was trained/captioned?yes, there was an interview with them, they're really focused on the open source part, they're really the good guys we've needed all this time, God bless them.https://podcasts.apple.com/au/podcast/the-researcher-to-founder-journey-and-the/id1740178076?i=1000665678592
>>101969043haha. cat eyes, eh. I could offer some husky eyes.>>101969100nope.>>101969123just don't copy pasta it, please.
>>101969159>nobody will find out I have an average size penis.OMG ANON ITS TRUE??
I get better results with loras in comfy, even using the q4 model, sometimes it wasnt working right in forge with the fp8 + clips + vae, as if the lora was sometimes workingheres a disney lora + miku: (forge)
>>101969205gonna try q6 or q8 now as I have 16gb, but it's already pretty good with q4.
arabian nights...
>>101969205protip:uninstall forge and keep a reforge around for compability. some of the flux loras very experimental. I've yet to try more than one at once. seeing nasty vram spikes too. looks good tho
how do you get average looking women out of Flux?
>>101969205>I get better results with loras in comfythe problem with loras in comfy is that it's unloading the whole model everytime you want to change a lora or change its strength, does it also do that in forge?
>>101969246I'm keeping both, I wanted forge mainly for quicker inpainting if I want to do that, plus each UI has pros and cons.
>>101969255forge does that too, comfy seems to be smoother for me so far, only occasionally do I get a long kmodel unloading spike with forge but it's mostly fine 99% of the time.
>>101969267plus this is really useful for monitoring resources:https://github.com/crystian/ComfyUI-Crystools
>>101969267I don't know why it does that, why can't it simply keep in memory the original weights that are about to be modified by the lora (20-80mb max) and then putting those weights back for each change
>>101969284yeah, I have a 4080 and if it kept the text encoder in RAM there should be plenty of space, it only does it periodically though. I guarantee these kinks will get ironed out as it's all very new. I like both comfy and forge, on a PC you have many different apps so theres no reason you have to exclusively use one.
>>101969297>I guarantee these kinks will get ironed out as it's all very new.it's been 2 years we've been using loras, and they are still using the retarded method of "destroy the modified model as a whole and reload a new one", that's ridiculous
I wouldn't, but many people on this thread would.
>>101969250painful a/b testing. try photographers in the clip_l. prompting with flux is a minefield. it's all there but one wrong token and you get slop>>101969257understandable. I got used to comfyUI inpainting eventually
>>101969318im just used to the gui as I started with auto1111, but I really like how you can do almost anything with nodes in comfy. also loading a workflow with an image is neat.
>>101969358also is there anything similar to civitai helper for comfy? most loras should be fine without specific instance prompts but some rely on them, is there a node or something that brings up civitai info/instance prompts per lora?
>>101969149Thanks! I'll check it out later
>>101969358most of us did. anus1111>comfyUI>oh look forge>back to comfyUI (because he actually is consistent) was my journey. can do stuff with comfy thats just not possible with forge or a1111 (ded anyway). UI improvements on their way, too.>>101969373rgthrees "power lora loader" node allows you to right click on a lora and receive detailed info straight from civitai. can even copy pasta trigger words , very helpful.
>An oil painting of a fantasy style portrait of an the angel Gabriel with wavy ginger hair, ethereal atmosphere, wearing silver full plate armor, holding a flaming sword, the brushstrokes are clearly visible. He is ready to beat the shit out of the devil and his demons.
>Can't game>Can't genWhat the FUCK do you guys do while training loras?
>>101969435Hold my dick in my hand while I wait for the next sample image.
>>101969435codewars.com
>>101969416>would. brushstrokes, cough. try "oil painting, brushstrokes" in clip:l, and then tinker with guidance and max_shift and base_shift>>101969435NO TRAIN JUST GEN
>>101969408perfect, that's exactly what I was looking for: it can even fetch civitai info (like civitai helper), so I can figure out what the instance prompt is if needed.at first I didnt see it in the manager but it must have been in one of the node packs I installed. double click and done.
>>101969342he looks like he'd be racist against aliens >>101969435why not do that while you're sleeping or working?
>>101969441It is quite the thrill slowly seeing it all come together in the sample images
>>101969469glad to be of service.
>>101969471:D
>>101969205Me on the right.
>>101969504very subtle.
>>101969551>we dont need lasers today, bitch
>>101969551yoda at a porno store
im not sure how but flux has a way of doing depth of field/perspective that makes photos look real, even if you do low quality as a prompt it has that low quality film look, or even do something like "CCTV" and it's in perspective, pretty cool.
>>101969622>boobs, you have?
>Maintain thread quality rentry.org/debo
pepe lora is great: also you can get any kind of pepe, for 2d just do "cartoon" or "comic".
12gb vramlet flux lora enjoyer back againtonights test will be :>lion+cosine >1024*1024 >8 dim, 8 alpha >1600 steps getting roughly 8 hour cook time, would probably be half that if I did 512*512 resohighest spike is:>9.4/12.0 GB vram >16.3/31.9 GB memtomorrow I will test how it turns out and report back
>>101969624I especially love how flux NUKES every gen with a brutal bokeh and depth of field. like, her fucking hand is out of focus. FUCK ME
Come and get a fresh baked loaf of...>>101969649>>101969649>>101969649
>>101969652when their txt2video goes live we're gonna have an open source luma.
>>101969632Boobs?
>>101969624I can't quite figure out the secret sauce for CCTV aesthetic
>>101969652>I especially love how flux NUKES every gen with a brutal bokeh and depth of field. like, her fucking hand is out of focus. FUCK MEthere's a way to remove thathttps://reddit.com/r/StableDiffusion/comments/1estj69/remove_the_blur_on_photos_with_tonemap_an/https://imgsli.com/Mjg3OTYw
>>101969504I'd watch this movie.
>>101969053SAUCE MEEEEEEEE
>>101971122TroonMix v7.7
>>101968360こもらし すて金けって。。。まちけてら。。。
>>101969665Any idea about how much vram video will require? I'm hoping it can get quanted like image models.
Pony user here, is Flux worth getting for spicy stuff?How does it compare when it comes to coherence and lewd content?
