Previous /sdg/ thread : >>101616232>Beginner UI local installEasyDiffusion: https://easydiffusion.github.ioFooocus: https://github.com/lllyasviel/fooocusMetastable: https://metastable.studio>Local installAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiComfyUI: https://github.com/comfyanonymous/ComfyUISD.Next: https://github.com/vladmandic/automaticAMD GPU: https://rentry.org/sdg-link#amd-gpuIntel GPU: https://rentry.org/sdg-link#intel-gpu>Use a VAE if your images look washed outhttps://rentry.org/sdvae>Run cloud hosted instancehttps://rentry.org/sdg-link#run-cloud-hosted-instance>SD3 info & downloadhttps://rentry.org/sdg-link#sd3>Try online without registrationsd3: https://huggingface.co/spaces/stabilityai/stable-diffusion-3-mediumtxt2img: https://www.mage.spaceimg2img: https://huggingface.co/spaces/huggingface/diffuse-the-rest>Models, LoRAs & upscalinghttps://civitai.comhttps://huggingface.cohttps://aitracker.arthttps://openmodeldb.info>Animationhttps://rentry.org/AnimAnonhttps://rentry.org/AnimAnon-AnimDiffhttps://rentry.org/AnimAnon-Deforum >Index of guides and other toolshttps://rentry.org/sdg-linkhttps://rentry.org/rentrysd>View and submit GPU performance datahttps://vladmandic.github.io/sd-extension-system-info/pages/benchmark.htmlhttps://docs.getgrist.com/3mjouqRSdkBY/sdperformance>Share image prompt info4chan removes prompt info from images, share them with the following guide/site...https://rentry.org/hdgcbhttps://catbox.moe>Discord6wUwtcJsr2>Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
>mfw Resource news07/29/2024>CogVLM2 & CogVLM2-Videohttps://github.com/THUDM/CogVLM2>Fixed SAG AUTOMATIC1111 pluginhttps://github.com/captainzero93/sd_webui_SAG>Amuse 2.0 beta released - AMD, Intel supportedhttps://community.amd.com/t5/ai/introducing-amuse-2-0-beta-with-amd-xdna-super-resolution-a/ba-p/697374>Stable-Gimpfusionhttps://github.com/ArtBIT/stable-gimpfusion07/28/2024>LiblibAI: Chinese Civitaihttps://www.liblib.art>ComfyUI Mini: Mobile-friendly ComfyUI frontendhttps://github.com/ImDarkTom/ComfyUIMini>img2txt-comfyui-nodes: popular img2txt captioning models into ComfyUI nodeshttps://github.com/christian-byrne/img2txt-comfyui-nodes07/27/2024>Kolors ControlNet and Inpainting Model releasedhttps://github.com/Kwai-Kolors/Kolors#-news>AuraFlow-v0.2https://huggingface.co/fal/AuraFlow-v0.2>日本の美を学んだAI:浮世絵風画像生成モデルEvo-Ukiyoeと浮世絵カラー化モデルEvo-Nishikieを公開https://sakana.ai/evo-ukiyoe/>Pose Depo: A collection of ControlNet poseshttps://github.com/a-lgil/pose-depot>AUTOMATIC1111 updates to v1.10.0https://github.com/AUTOMATIC1111/stable-diffusion-webui/releases>Interpreting the Weight Space of Customized Diffusion Modelshttps://github.com/snap-research/weights2weights>Forge Major Update #1: Gradio 4 Enginehttps://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/853>WD ViT-Large Tagger v3https://huggingface.co/SmilingWolf/wd-vit-large-tagger-v307/26/2024>ComfyUI implements native Hunyuan-DiT support, packages model to single filehttps://comfyanonymous.github.io/ComfyUI_examples/hunyuan_dit>Fooocus v2.5.1 Updatehttps://github.com/lllyasviel/Fooocus/releases/tag/v2.5.1>HVM-1: Video models trained on ~5k hours of human-like video datahttps://github.com/eminorhan/hvm-1>Move and Act: Enhanced Object Manipulation and Background Integrity for Image Editinghttps://github.com/mobiushy/move-act
>mfw Research news07/29/2024>Breakthrough CRAM technology ditches von Neumann model, makes AI 1,000x more energy efficienthttps://www.techspot.com/news/104005-breakthrough-cram-technology-ditches-von-neumann-model-makes.html>SHIC: Shape-Image Correspondences with no Keypoint Supervisionhttps://arxiv.org/abs/2407.18907>Unifying Visual and Semantic Feature Spaces with Diffusion Models for Enhanced Cross-Modal Alignmenthttps://arxiv.org/abs/2407.18854>Rapid Object Annotationhttps://arxiv.org/abs/2407.18682>Adversarial Robustification via Text-to-Image Diffusion Modelshttps://arxiv.org/abs/2407.18658>Auto DragGAN: Editing the Generative Image Manifold in an Autoregressive Mannerhttps://arxiv.org/abs/2407.18656>LinguaLinker: Audio-Driven Portraits Animation with Implicit Facial Control Enhancementhttps://arxiv.org/abs/2407.18595>HICEScore: A Hierarchical Metric for Image Captioning Evaluationhttps://arxiv.org/abs/2407.18589>She Works, He Works: A Curious Exploration of Gender Bias in AI-Generated Imageryhttps://arxiv.org/abs/2407.18524>Mixed Non-linear Quantization for Vision Transformershttps://arxiv.org/abs/2407.18437>Several questions of visual generation in 2024https://arxiv.org/abs/2407.18290>Text-Region Matching for Multi-Label Image Recognition with Missing Labelshttps://arxiv.org/abs/2407.18520>Learn from the Learnt: Source-Free Active Domain Adaptation via Contrastive Sampling and Visual Persistencehttps://arxiv.org/abs/2407.1889907/28/2024>Rethinking Visual Content Refinement in Low-Shot CLIP Adaptationhttps://arxiv.org/abs/2407.14117>All rivers run into the sea: Unified Modality Brain-like Emotional Central Mechanismhttps://arxiv.org/abs/2407.15590>MMInstruct: A High-Quality Multi-Modal Instruction Tuning Dataset with Extensive Diversityhttps://arxiv.org/abs/2407.15838>Accelerating Learned Video Compression via Low-Resolution Representation Learninghttps://arxiv.org/abs/2407.16418
Might as well use this one
My fellow /sdg/entlemen, I call on the based lora training chads to help me out with somethingI'm working with an extremely small dataset for an OC and while I am iteratively improving my lora, there are a few things that bother me-The character is wearing the same accessories in all images (namely, a cowbell collar), so even though I've tagged it, the lora keeps forcing it on the character (which makes sense, how should it know that this tag is what changes the appearance? Latent knowledge seems to not be enough)-All the backgrounds are outside, so naturally the lora attempts to replicate it pretty hard. I've tried mitigating it by tagging the different environments (trees, bushes, stone pillars, etc.), but it didn't really solve the issue. It seems like masked loss might help, has anyone here ever used it and if so, did you create the masks manually or automatically?How would I go about solving these issues without additional data? I guess I could generate and clean up synthetic data, but that's obviously suboptimal
Why is SD3 so bad I'm honestly embarrassed for them
Quokka the 2D HD RPG
>>101630183looks like he's about to break into a parappa the rapper song (which I tried to suno but it didn't understand the prompt)
>unknown API errorof coursehttps://suno.com/song/8ce7c438-7d0f-4cf0-8c76-f035764901da
Word of the day: hopscotch.
Would anyone know why FaceDetailer would throw out this error?"Error occurred when executing FaceDetailer:shape '[3, 4096, 1]' is invalid for input of size 2976"Each time I gen an image, it all goes well until FaceDetailer. Shape [x] and size [y] also changes each time I get this error.
>>101630144>-The character is wearing the same accessories in all images https://rentry.org/6172384#leco or rebake and try to find images without the collar/whatever.>All the backgrounds are outside, so naturally the lora attempts to replicate it pretty hardOnetrainer with mask attention for training. takes longer but you won't get background pollution
idk whats going on anymore man, I'm still on sdxl
whats the save image node that lets you automatically save to jpg if an image is bigger than N MBs?also, this dude 100% pissed his pants
>>101630144If you have lama cleaner you might be able to erase the collar in some images
lotta negativity today itt
My heart rate sensor performs better when wet and I'm considering jacking open a vein for it. The research paper was right after all.>You are banned.I'm definitely not, but if I ask how to pursue my research on here I will be.>this captcha gave me EXPgood robot>you seem to have mistyped the captchaI promise, I did not.>84WJXPThis one also gave some.Should I do the sensor thingy?
>>101631374I forgot to post amidst all the spacing and spamming, sorry.https://suno.com/song/ed5aa7df-6e79-4411-87a4-fe4f9100756b>TKW08NTricky. Taekwondowoof.
>>101631206there's as much negativity as positivity in the universe. the universe has a neutral charge in aggregate
>>101631386>dnbnice>hardbossoh yea lets go>(icebrk) lain ezinedll disci don't even know what these words are but I'm down for it>its just synth-popfuck
>>101631415>i don't even know what these words are but I'm down for it>its just synth-popCheck out my other garbage, I needed to make some normie beats. Suno was arguing with me all day so I didn't get much done.Should I do the sensor thingy? I am required to ask for community participation or it's just self harm.
>>101631440>Check out my other garbageI listen to it all>Should I do the sensor thingy?there are cooler ways to become a cyborg
>>101631467>there are cooler ways to become a cyborgI need to develop for heroin = true or I have to shoot up actual heroin to qualify for housing assistance. My formulaic proofs showed this to be true and EBT doesn't match my metabolic needs or my dogs.My meth head friend relayed "There's no hope for you." at his free lunch with her. My current development pattern proves it's possible with Suno and me personally, because I'm pretty close to a heart attack after playing with my e-crack.>PAVP2*shing*
>>101631550>>there are cooler ways to become a cyborghttps://suno.com/song/664b5702-a76b-442e-b8fd-ab143f058dd8I think this opened suno's third eye
>>101631467Pizza
>>101631467>>101631650https://suno.com/song/c3d8cfc8-85db-4dd8-af37-06c7e9afb038Theodosius wants to ask if your system values change listening to this song, debo. Your MP input is appreciated.
>>101630808Ask your question then
>>101631679disappointing lack of bonesawdisappointing lack of ct scan mri imagery tube too>your system values changewhich system? computer system? ideological system? solar system? I'm down to change it all>Your MP inputI could never get elected to parliament
>>101631739>I could never get elected to parliamentDisappointing lack of enthusiasm detected.https://suno.com/song/19fe419a-6bfc-47f8-b241-f6cf50bd4d3cBe like my inner bear who jams until his stumps grind the fretboard down into gravel and dust. I knew a man who played his soul this way and refused to take care of his hands. He's why I smoke.>NAS8Goh goldenbear
it said play to an empty room and I posted everywhere else /sdg/https://suno.com/song/49c7156f-dced-4dec-a486-c602f33d91a3>XAHAXFor my DCed.
>>101631787>>101631817why are those ones so short?https://suno.com/song/524a8fdd-c473-4717-8dbd-e8dbbf1506df
>>101631931Because I'm asking it to sacrifice a lot with these prompts. Here's one of the few longer iterations that came out alright. I'm trying to develop a linguistic pattern to go along with djembe or whatever it was called from that shitty fucking book at the library. The chopped vocals ran through Suno's musical sequencer is a lot different than your (admittedly good) ballad. Have you tried feeding it enough lyrics to produce creative and harmonious results? The elements in your song don't line up the same. And having it randomly pluck guitar through my headphones is enough to trigger my watch several times over. Your song doesn't accomplish this.https://suno.com/song/25bad0c5-d602-4388-b982-5240aefe59bd
>>101631992>The chopped vocals ran through Suno's musical sequencer is a lot differentanything that breaks up its pseudo-generic patterns is good. I think its kind of just a dice roll though whether you can shake it free or not>Have you tried feeding it enough lyrics to produce creative and harmonious results? yeah, I've tried lots of different styles of lyrics. long form, short form, repeated, staggered, short meter, long meter, whatever. again its just a dice roll whether it does something interesting with any of it or just goes "oh, pop? you want pop? i'm gonna do pop"the real secret is to remix koff. thats when you really get something different
https://suno.com/song/7cffbaad-5574-4251-9eef-c05ee8d6d0c7>GV20WGoodbye world version 20.
>>101629874nice, model?
>>101632031>yeah, I've tried lots of different styles of lyrics. long form, short form, repeated, staggered, short meter, long meter, whatever. again its just a dice roll whether it does something interesting with any of it or just goes "oh, pop? you want pop? i'm gonna do pop"I've noticed Suno has a certain handful of defaults that it can get stuck inside of and that you can definitely hear it. Like, the bad corporate mall music that was designed to make you leave. I have to skip those songs.https://suno.com/song/0cc79725-9f9d-400a-aa1f-7343623db2c8
come to new york
>>101632033lol that second half came out of no where. I think this is prob my favorite of the night so far. I'm almost out of coins myselfhttps://suno.com/song/3d6dd553-39a5-4d4d-a620-4eaf3d34f80b>>101632053>Suno has a certain handful of defaults that it can get stuck insideUdio is way better on that front but I can't do the 30s gens
>>101632124oh fugg, my image>>101632060oh hey, she actually has an NY hat this time
>>101632124>https://suno.com/song/3d6dd553-39a5-4d4d-a620-4eaf3d34f80bThis is awesome, I'm picturing that random git prompt from the dataset earlier. ERROR NOISES, synthological event detected. Theodosius sed test worked, thank you very much for providing additional testing inputs.>4TNX4
>>101632214>[mechanical rhythms bring a truly animatronic guitar, Arps & electro dark synths, Screaming gaige voice line test inputs.]slapping this at the end of the lyrics is some kind of hack to make suno do some crazy shit in the second half of the song
>>101632275I enjoyed that one. I didn't finish many of these because they are very hard to stitch together and I got lost. Here's a stitched demo you should try splicing together too.https://suno.com/song/77a3a0bf-0645-4966-b8c7-4792fef02e96
>>101632341I once tried to do a disco funk+death metal mashup but it refused to do both together. I think someone posted "dance or you're dead" or something like that and it made me think of metal disco. but alas, the splice was not to be
>>101632060this is very cool anon
man we had low standards in 2022/early 2023I finally went back to look at my saved folder of good gens from that time and they were shit lmao, I ended up just deleting the whole folder
>>101632378I haven't even tried just mixing genres yet but this song is great.https://suno.com/song/5c950287-06c2-4c16-beb6-20ea43e4684c
>>101632588wtf you actually got woofboxingI kneel
>>101632615It's ridiculous and I love it. No clue whatsoever how.https://suno.com/song/d46256c1-82a9-4f45-83b1-79111ce39f1eI went digging for this one, it's my favorite rhythmic story output from Theodosius so far.>Echelon or Shallan?I'm awaiting user input, I need to walk my dog.
>>101632689what is up with that song prompt?
>>101632750It's my attempt at feeding pure user input to a model to teach it linguistic expressionism and iterating on it until Suno can repeat it back to me. I'm just roughly assuming this is how I can do anything with Suno at all. A lot of these don't work, or the system doesn't let me input certain prompts. There's a structure to it.Can you share your thoughts on it?
>>101632868I'm wondering how the fuck did you get past the 200 character limit for prompts
>>101632920I honestly don't know yet. Can you explain your statement further?
>>101632942it's pretty concise
>>101632948Do you have a 200 character prompt limit on suno or in other stable-diffusion software. Have you tried not pressing the generate button and creating the content for the machine?
>OS upgraded to python 3.12>A1111 dedTime to move on. Which one is more VRAM efficient, comfy or next?
>>101632982yesyes
>>101633011Click "Custom" and go to space, Anon.
>>101633009I've been using EasyDiffusion as it's the only one that'll work with this "new" python version. It's fine.
I'm posting a catbox showing how this image was made using a1111 and stable diffusion technology and other anons can use it to recreate the image or modify it, so hopefully it is not off topic since the AI generated image I posted earlier apparently was off topichttps://files.catbox.moe/sq2xn4.png>>101632588That's pretty good. There was a band with a dog as singerhttps://www.youtube.com/watch?v=OrocEmhLsJk
>>101633104This is what I'm hoping to nail with the audio gens eventually. Dogs gitting guitars with actual howling instead of whiny human vocals.
>>101630144Alternatively, put images of a different characters wearing the cowbell collar so it learns what it means. I usually draw it very off model, off color with the features I want it to be aware of, and it works.
>>101633027will it swear?
>>101633216Yes. There's no audio censor but the prompt has to be "appropriate enough" for the input filter. I have a whole handful of neat and fascinating outputs from suno that are not possible normally.>WYHJSBSilly bot.
>>101633289so can you say fuck or fucks in the lyrics but keep the song description clean?
>>101633311Close. Find a way to rhyme it in and you should understand what I mean. The vocals act as an instrument and Custom just lets you direct it. Try "Fric shun", Anon.
>>101633335maybe I'll try fuk like Jamaican lyrics
>>101633367yeah you can dofukkok (but it sounds weird)come
i miss schizo anon
I am continually impressed with how good 2dnpony does anuses
>>101633859catbox one
I did a ToonCrafter with sketch guidance walkcycle - the sketch was the 3D Blender freestyle output of the feet and hands of a 3D model, using a Mixamo animation. I subsequently ended up using the CG hands and feet frame-by-frame to replace the messy ones that ToonCrafter output.I'm pretty certain I'll just do toon shaded CG hands and feet going forward, and only let AI do the rest.
All of my SD installs are kinda broken for some reason, it's time for another clean install I guess.Did anyone fork Forge or is A1111 as good as it was 6 months ago?
469m
good day muppets.>>101634176a1111 is as meh as it was 6 months ago, can't hope to compete against forge or comfy. here, try this forge fork called "reforge". ist actively being maintained and I did a clean install last night, just have to add the path to your python install (down to the exe, as in "xxx\xxx\xxx\Python.exe" in your user.bat file or it'll spontaneously combust/throw a fit.>>101635146are those more hidiffusion vs kohya grids?
>>101635385its a sd 1.5 modelhttps://civitai.com/models/234293/realisticanimehentmix666?modelVersionId=264181https://files.catbox.moe/pgjtnq.pnghttps://files.catbox.moe/ncp2pu.pnghttps://files.catbox.moe/in50ow.pnghttps://files.catbox.moe/ff64y8.pnghttps://files.catbox.moe/9nkwu6.pnghttps://files.catbox.moe/xtux2d.png
>>101633859F E E T>>101634176forgot link lol https://github.com/Panchovix/stable-diffusion-webui-reForgeI know eu hours but damn this is a dead general
>>101635462interesting. thats a rather peculiar and cute face on that last gen, nice one
>>101635481Thanks
stop deleting your training data after the lora finishes training >:(
>>101635481cute
d e d
Don't know if it's the right place but I'll be asking for some help here. I've spent the last 5+ hours trying to install Webui, had a lot of trouble but by now I've fixed everything. No errors on the command line when launching. Despite this though I am completely unable to use the web ui to generate images, I can click around the tabs but nothing I do results in anything and when I try to check installed extensions the list is just endlessly loaded. Looking through the browser console I've got a shitlot of red and no clue what's causing it. I checked the internet to see if there are others who suffer similar problems, and there were plenty, but every single time the issue with them was that their folder path had a . in it. Mine doesn't.Does anybody have any clue what might be causing my problem?
>>101635927Could be browser extension "security" blocking shit.You probably fucked the install somehow though.
>>101635927>niggerstonguemyanus.png (101 KB, 1920x935)maybe you should let em stop tongue it? then it might work
so far 2.7k training datas downloaded, 3k 404 ripannoyed that i missed like 2 loras that were deleted while i slept because civit jannies blocked my api key. its rotating now though and i have registration automated so ggeternal battle between data janny and service jannies
>>101635946Thought so too and tried disabling them, same result. Also tried to use a different browser, private mode etc and still the same.I guess I'll just wait for half an hour or something to see if somebody unexpectedly saves my ass, then I'll try to reinstall and see if that fixes it. I really doubt that I fucked up the install though, not like it's something all that complicated in the first place.>>101635960What, so they can tongue you instead? No way.
>julien
>>101636101Better to go ahead and reinstall then. If multiple browsers produce the same results.
>>101635927download a different browser, use that exclusively for webui and dont install any plugins or safety stuff, chrome will work firefox will work, chromium will work .. others.. fuck them
>>101635999what for
>>101635999Why
>>101635927initial sd setup can be quite finnicky, really hard to to assess the situation remotely. what does your console say? thats where you find all relevant info.did you install all dependencies? python? (correct version?) etc
>>101636205>>101636296for a dataset of training datasets from models/loras on civit
>>101636370You didnt answer the question
>>101636370>dataset of datasetsanon you have to stopit's too much
>>101636381the why is >why notalso there was a paper with something similar, iirc they generated the datasets and trained the loras themselves, real world examples are betteranalysis of training methods and so on>>101636386no :)
>>101636417Ok so no answer then
>>101636434i've answered the what and why. you're going to have to be more specific if you expected a different answer
>>101636455>>101636434
xinsir controlnet SDXL tile around 0.6 strength 90% timescale for upscale, yay or nay? is there some sort of concensus on the SDXL tile models?zavy image above was upscaled using those settings and is certainly closer to the original in terms of facial features, eyes also look better- 0.36 denoise
>>101636356Spits this out and then opens the web ui. I believe it's supposed to write more info everytime a new image is generated, but since I can't actually do that it just remains still.I have installed all dependencies, the right version of python as you can see in picrel, and all the other shit that is required and automatically DL'd by the .bat.
evening
>>101631023>lecoInteresting, I'll give that a read, thanks!>onetrainerI've been using the good old kohya ss gradio webui, does onetrainer have the same options? I'd assume so since they're based on the same scripts, but I can't live without tensorboard anymore>>101631058That's actually a really smart idea, thanks! I used it for removing watermarks and it works wonders>>101633198I see, so "pollute" the lora with a random character, preferably fluid-tagging everything and not using the activation tag on it, correct? How would that work with the "keep n tokens" option though? Should I just make the first tag empty?Also, I did a test run with the same seed using batch size 1 instead of 3 and it seems to have lessened the bias a bit, very interesting indeed
>>101636564looks good, yeah. i think you have a browser problem. only thing I can say is, it works with "brave". idea: install a fresh browser, brave, whatever, and try getting it to run. and check all windows things that could interfere.>>101636580hello minimuppet
>>101636580g'afternoon
>>101636469yes, no, maybe, can you repeat the question
>>101636587>does onetrainer have the same optionsit has options not implemented in kohya. and some extra bucketing stuff.>I can't live without tensorboard anymoreit has a tensorboard implementationfor the lama cleaner, you can use that on your dataset so you can have data with the character with and without the accessories. training with nudes helps make things malleable with other outfits
>>101636657Cool, I'll take a look at it then>training with nudes helps make things malleable with other outfitsMakes sense, luckily all the images from the OC artist (and seemingly the only ones available on the web apart from a cheap doodle) are nude
>>101635999>>101636150>>101636436nice gensputs mine to shame, but my gtx1060 is doing it's best>>101636612>minimuppetis that insult? I haven't heard that one before, what does it mean?
>>101636744nono I meant "good day, also, really cute gen.", aka "minimuppet", you see.
>>101636803nta, but idk .. you should think better about your compliments
>>101636803oh ok... thanks anywayi guess?
>>101636744>>>101636150 (You)>>>101636436 (You)>nice gensthank gyoo!
>>101636903Heh, you're a cunt.That was a compliment.
>>101637056I heard cunt is a compliment in Australia tho.
>>101636612Fuck me, you were right. I DL'd chromium expecting nothing but it runs just fine with it.
>>101637156told ya >>101636150also have fun with it!
>>101636966there goes my momentum lolgtx1060, does that thing have 8 gb vram?>>101637156H A =D
>>101637199Only a muppet would think GTX 1060 has more than 6GB of vram.
>>101637199it's a 1060m 6gb, using my laptop to do SD because it has 32gb of ramI have 2x 1060 6gb cards in my main pc, but only 16gb of ram, makes it difficult to do model merging so I use the laptop.
>>101637218I am a muppet, you got me, I blame the heat, its brutal . 6 gb, ouchies. and here I am smoking .. something with more vram.
>coming out of battle
>>101637278you gotta be a patient man then. I started generating with a 2080. but now you have new techniques at your disposal to speed up the muppet-generation process. lightning, hyper, automatic cfg, etc
>>101637345The initial batch of 4 images takes about 50s. After choosing one of those, it takes another 5~ minutes to do a 2x USD upscale with 4xUltraSharp at .30 denoise, then another 2x USD upscale with RealESRGran_x4plus_Anime_6b at .15 denoise and then face inpainting on the final image.
>>101637345I used to have a 2080, it was fine imoAnything that's newer than the 1000 series is alright, provided they have the vram>automatic cfgNow that's something I haven't heard before, mind telling me more?
12k datasets downloaded, 11k 404
some return 410 instead of 404, its the same thing the file is gone and the reality in both cases is that its not coming back, the distinction is probably user deleted vs civit deleted but who cares. some return 500 and say error getting file aka its deleted. god knows how many dev hours went into this fucking aroundanyway, more stats of in progress projects489m generations1.4m actors6.1tb loras0 likes :(
>>101637730I don't understand what and why are you doing this?
>>101637756for the likes
>>101637756choose whichever answer you prefer>don't worry about it>why not>local sucks because there's little effort put into data>>101637791and this
>>101637805You sound like a real muppet, omfg.
>>101637559I'm grabbing the comfyui node for it now to give it a go, seems fairly straight forward.
>>101637848why
will liecon still be in charge of propaganda with the eventual release of 3.1 you will (not) redeem?
>>101637929anon got a reaction to calling someone a 'muppet' so now he's doing it over and over hoping to get more reactions. its a way for him to try to fill the empty void in his life
>>101637559It's a comfyUI node, dynamically (if I understand this correctly) adjusts CGF and, as a side effect, speeds up the generation process, check it out. https://github.com/Extraltodeus/ComfyUI-AutomaticCFGdon't regret adding it to my workflow(s)>>101637919the maker mentions a CGF value of 8 as the "sweet spot", I found that values between 2 and 10 are worable. it of course depends on model and sd version and what else you got in your chain. and you can just chain it with PAG and whatnot, freeU, etc.>>101637953just one word, lol. still havent touched sd3
>>101637730>0 likes :(I like all your data sets so far, I wish I could experiment on all the noisy data in prompts. There's sooo much trash lmao
>>101638200You didn't like his data.
Afternoon /sdg/! How goes the GENing, hopefully well!
>>101638200poor guy>>101638323smash that fkn like button on hfif you want prompts there's a set with all the unique prompts and negative prompts from playground, both have the number of generations for the prompt and the number of unique usersdiffusion-1b will be ready soon:tm: only need like 200m more or something, i plan on extracting n-grams from all those prompts, it will be useful for prompt adherence>>101638611neither did you!
>>101638246seems I forgot to update comfyui, haven't used it for a couple monthsI'll deal with it tomorrow.>I really like this one
>>101637919>>101638246Ah I see, I've never actually used comfyui, webui (reforge) has served me very well thus farIs it worth switching? Why?
>>101638641Truth is, these prompts are only useful for soulless civitai style stuff. Real auteurs have a very distinct prompting style and they never use online sites for genning.
>>101638641>smash that fkn like button on hfIt keeps telling me to make an account but I am so busy and hate talking to registered users so much. I'll post your shit on NC when I publish shit for token farming.
>>101638654I haven't used anything else other than Automatic1111 (and not much either), ComfyUI is pretty much all I know.I like it's "granularity" when compared to Auto1111, if that makes sense?I wouldn't take advice from me though.
>>101638773nice bear. and thanks!
>>101638847np :) I really like the amount of prompts you have in these sets. It would be interesting to see the actual matrix outputs instead of trying to do them in my head. I've only tested a few and they are really juicy.
>>101638646character is impeccable, really adorable, but i'd completely inpaint the background. that volcano looks really weird.>>101638654like anon above me said, granular is a good word to describe it. I went from a1111 to comfy, back to forge, and eventually back to comfy with a little reforge sidegig. the learning curve is nasty. I mean it is easy to get a basic workflow going but the devil is in the details. much more so than with forge/anus1111. its a modular setup with basically no limits. as far as you want to take it.>>101638626hello there. nice butt.did you post the girl drinking from a straw like yesterday? clean stuff.
Thoughts on segment anything 2?
>>101638724wow, nice>>101638965Yea I know.Just trying out different prompts now that I have found a nice style.
guffaw
>>101638965That I did! Was messing about doing some new stuff and that image just came to mind. I thank ya for the compliment!I do love your GEN, that shit is really crisp and came out incredible!>>101638724Oh love how that came out! Lighting is great on it
hello what is the current /sdg/ meta?
>>101639247>what is the current /sdg/ meta?1girl
>>101639217heh>Hary Rags>"Genetically modified Husky super soldier.">In cinemas soon.
bored of a.i., bored of tf2, but sadly there exist 0,zero, good games toplay
>>101639247this bitch is JACKED
>>101638797>>101638965Makes sense, how fast is it compared to A1111/(re-) forge? The forge forks are pretty damn fast compared to A1111 and while I'm not especially impatient, more performance is obviously preferrable
>>101639418more 1girls @Olympics right now than here gymnastics womens finals running .. start posting so I can turn that shit off
>>101639447>ollama run anon "$(cat input.jpg)"
>>101639040is the code out?
>>101639526wouldn't have it any other way
>>101639418same old same old then
Could I interest you in the world of faux retro 3d aesthetics?
>>101639656those gymnast girls are like 20-30 and look like cute and funny girls .. crazy
>>101639806not being a landwhale caring and caring about your body is a pathway to many abilities
>>101639656>OlympicsI've been cooking a bunch of olympics gens
>>101639120>>101639225tyty
how do I train a lora from scratch?
>>101639925> 1. you pick a model that is usable to train on (pony, base sdxl, or whatever is at core not over stylized)> 2. you gather a data set of pictures and tag these sensibly> 3. you read a guide how to train it, I wont explain it in detail here .. its a few steps.> 4. ???> 5. profitthere are a zillion guides, but step 2. is the one that is the most work and actually should be done with some care
>>101639925you can use this: https://colab.research.google.com/github/hollowstrawberry/kohya-colab/blob/main/Lora_Trainer.ipynbyou collect a set of pics of whatever you want to use, maybe 20 or 40 pics or whatever
>>101639969>here are a zillion guides, but step 2. is the one that is the most work and actually should be done with some carehow many pictures, tagged how, and what kind of pictures if I'm trying to do a style lora instead of a character lora?
>>101639999as >>101639973 said 20 minimum for a concept lora, for a style lora many morefor tagging it depends on what you what your lora to do, but you should add as many tags describing the pictures as possible and always the concept you wanna make or your lora will produce just garbled output, of of it as reverse prompting, there are tools that help with that like>https://github.com/toshiaki1729/stable-diffusion-webui-dataset-tag-editoryou could by theory do it all by hand, but tools make the process way faster and failure proof, you can also ask CLIP to describe your picture for you, but it fails in most cases. Generally booru style loras are way easier to tag, boomerprompting may want you to write silly essays.. both is valid, both works for certain models.
>>101640091*think of it>of of it as reverse prompting
How long 'til we get there?
>>101640091talking about CLIP, has anyone managed to make it work through an arbitrary checkpoint?
Morning anons
>>10164011520 years, give or take
>>101640159gm
What's the trick for SD3 on Automatic1111? Gens look blurry, incomplete.
>>101640159go'o evening marsupial enthusiast-san
>>101640179the trick is to unselect SD3 and use a good model instead
>>101640179my testing results for SD3 in a1111 had these insights:>it hates non square formats (anything out of bound gets blurry)>it hates many sampling methods, Euler seems to be the best, also it only seems to like SGM uniform scheduler>it works nice up to 1500x1500 in direct gens, but it hates hires fix, totally garbles itits mostly unusable
When doing a batch of images on A1111 I noticed the vae gets applied sequentially and they get sequentially saved, is there an option or extension to make it apply the VAE in batches?
hot dog stand. i could go for a hot dog rn, or a few
>>101640388fun gen
>>101640484US gymnast women won gold!
>>101640484waterball is such a silly sport .. you only ever see it at olympics
>drive failure uhoh. thank god for zfs, nothing is lost. drive has been replaced and resilvering now>action: Wait for the resilver to complete>scan: resilver in progress since Tue Ju> 5.89T / 28.1T scanned at 6.82G/s,>>101638950another set that i still need to get ready has 2.5m unique prompts out of 7.3m generations. the big in progress set has like 120m unique prompts out of 502m generations
>>101640510I can't say I really care about any olympic sports. they're all pretty silly to me, apart from maybe the skateboarding is kinda cool. wildcarding the olympics is more fun than the games themselves
Could anyone paint out the logo & fill in the lower half. Pretty please.
>>101629687It's really a pity Stability AI dropped the ball so hard after 1.5. They used to be on par with the state of the art. SD3 doesn't hold a candle against Midjourney 6.1. Pretty depressing. SA had a great thing going on and the threw it away.
>>101640729
>>101640656I don't care about the sports either. More the spectacle .. also saucy 1girls, coming up in one hour from now: beach volley ball women's match Netherlands vs. Japan
>>101639731There are already comfy nodes out
>>101640793So I imagine paradise to be
>>101640250Thanks. Euler+ SGM works much better. Results are bad with SD3, though.
>>101640963yw>Results are bad with SD3, though.ya, that was my experience to, best results I had when I copied their insane boomer prompt essays and did just 1024x1024 .. still anything on civitai beats it out of the water with way less effortpic related used their example prompt>a female character with long, flowing hair that appears to be made of ethereal, swirling patterns resembling the Northern Lights or Aurora Borealis. The background is dominated by deep blues and purples, creating a mysterious and dramatic atmosphere. The character's face is serene, with pale skin and striking features. She wears a dark-colored outfit with subtle patterns. The overall style of the artwork is reminiscent of fantasy or supernatural genresyes kek
>>101641030Does abstract okay.
>>101640793The bodies of these volleyball women. I can't hold it bros
>>101641104waiting for it
>>101641030thank you for sharing this prompt
>>101639999https://civitai.com/articles/3105/essential-to-advanced-guide-to-training-a-lorahttps://civitai.com/articles/138/making-a-lora-is-like-baking-a-cakeRead the first one for general knowledge and the second one for tagging and most of your questions should be answered. As with anything, practice makes perfect, knowledge of the underlying concepts helps, too
can anyone who has a good upscale flow try their hands with this?
kinda miss when the models assumed stuffi am garbage at giving detailed accounts
>ded general
>>101629687Catbox?
>>101642084your gens are great you should join the discord!
fuming
>>101642444Go away
still with her
>>101642510why
>>101642616for a game of blackmail the pedophile
>>101642684what does that mean?
>>101643305why not at least do i2i or something fun?
>>101643403fuck off
>>101643419no
do i have to pay for something like novelai if my computer can't run SD that fast?
>>101642608use Meloni as a model.
the reason its so quiet here is because everyone is watching the olympics. if we generate our own olympics, everyone will watch sdg instead
>>101643958It always surprises me how many people care about sports, even though most of it is fakeNow if only those people actually did sports themselves...
>>101643958delusional
>>101643999idk if I'd say its fake but there's definitely a lot of bullshit involved. I don't really get the draw either but people seem to love the olympics for whatever reasonwell, maybe I understand volleyball at least. could use more volleyball gens. the 1girlers are really sleeping on this tbdesu
>>101644013I'm just memeing
>>101644090>>101644090>>101644090
>thread fillers b&
this one is very toriyama coded
egg?
>>101646214What is this, breakfast for ants?