Previous /sdg/ thread : >>100134744>Beginner UI local installFooocus: https://github.com/lllyasviel/fooocusEasyDiffusion: https://easydiffusion.github.io>Local installAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiComfyUI (Node-based): https://rentry.org/comfyuiAMD 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>Auto1111 forksForge: https://github.com/lllyasviel/stable-diffusion-webui-forgeAnapnoe UX: https://github.com/anapnoe/stable-diffusion-webui-uxVladmandic: https://github.com/vladmandic/automatic>Run cloud hosted instancehttps://rentry.org/sdg-link#run-cloud-hosted-instance>Try online without registrationtxt2img: https://www.mage.spaceimg2img: https://huggingface.co/spaces/huggingface/diffuse-the-restInpainting: https://huggingface.co/spaces/fffiloni/stable-diffusion-inpaintingpixart: https://huggingface.co/spaces/PixArt-alpha/PixArt-Sigma>Models, LoRAs & embeddingshttps://civitai.comhttps://huggingface.cohttps://rentry.org/embeddings>Animationhttps://rentry.org/AnimAnonhttps://rentry.org/AnimAnon-AnimDiffhttps://rentry.org/AnimAnon-Deforum >SDXL info & downloadhttps://rentry.org/sdg-link#sdxl>Index of guides and other toolshttps://codeberg.org/tekakutli/neuralnomiconhttps://rentry.org/sdg-linkhttps://rentry.org/rentrysd>View and submit GPU performance datahttps://docs.getgrist.com/3mjouqRSdkBY/sdperformancehttps://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html>Share image prompt info4chan removes prompt info from images, share them with the following guide/site...https://rentry.org/hdgcbhttps://catbox.moe>Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdgOfficial: discord.gg/stablediffusion
>Dude how are your LORAs so good?>idk dude, I just throw a bunch of images into a folder and hit train.
>mfw Resource news04/22/2024>MoMA: Multimodal LLM Adapter for Fast Personalized Image Generationhttps://github.com/bytedance/MoMA/tree/main>LLaMa3 Stable-diffusion prompt makerhttps://ollama.com/impactframes/llama3_ifai_sd_prompt_mkr_q4km>PhysDreamer: Physics-Based Interaction with 3D Objects via Video Generationhttps://physdreamer.github.io/>Training-Free Painterly Image Harmonization Using Diffusion Modelhttps://github.com/BlueDyee/TF-GPH>TV100: A TV Series Dataset that Pre-Trained CLIP Has Not Seenhttps://tv-100.github.io/>Hyper-SD: Trajectory Segmented Consistency Model for Efficient Image Synthesishttps://hyper-sd.github.io/04/21/2024>FlashFace Inference Code Releasedhttps://github.com/ali-vilab/FlashFace>ComfyUI MagickWand: Proper implementation of ImageMagickhttps://github.com/Fannovel16/ComfyUI-MagickWand>Moving Object Segmentation: All You Need Is SAM (and Flow)https://www.robots.ox.ac.uk/~vgg/research/flowsam/>Image Effect Scheduler Node Set for ComfyUIhttps://github.com/hannahunter88/anodes/>ComfyUI-Tripo: Generate 3D models using the Tripo APIhttps://github.com/VAST-AI-Research/ComfyUI-Tripo>Conditional Prototype Rectification Prompt Learninghttps://arxiv.org/abs/2404.0987204/20/2024>Basic Stable Diffusion API GUIhttps://github.com/ThioJoe/BasicStabilityAPI-GUI/>IPAdapter Advanced Weighting support added to sd-webui-controlnethttps://github.com/Mikubill/sd-webui-controlnet/discussions/277004/19/2024>Customizing Text-to-Image Diffusion with Camera Viewpoint Controlhttps://customdiffusion360.github.io/>StyleBooth: Image Style Editing with Multimodal Instructionhttps://ali-vilab.github.io/stylebooth-page/>Sketch-guided Image Inpainting with Partial Discrete Diffusion Process https://github.com/vl2g/Sketch-Inpainting>ComfyUI ImageMagick: Image processing powered by ImageMagickhttps://github.com/jtydhr88/ComfyUI-ImageMagick
>mfw Research news04/22/2024>Analysis of Classifier-Free Guidance Weight Schedulershttps://arxiv.org/abs/2404.13040>Neural Flow Diffusion Models: Learnable Forward Process for Improved Diffusion Modellinghttps://arxiv.org/abs/2404.12940>Learn2Talk: 3D Talking Face Learns from 2D Talking Facehttps://arxiv.org/abs/2404.12888>MCM: Multi-condition Motion Synthesis Frameworkhttps://arxiv.org/abs/2404.12886>GenVideo: One-shot Target-image and Shape Aware Video Editing using T2I Diffusion Modelshttps://arxiv.org/abs/2404.12541>Compositional Neural Textureshttps://arxiv.org/abs/2404.12509>Global Counterfactual Directionshttps://arxiv.org/abs/2404.12488>MoVA: Adapting Mixture of Vision Experts to Multimodal Contexthttps://arxiv.org/abs/2404.1304604/21/2024>Prompt-Driven Feature Diffusion for Open-World Semi-Supervised Learninghttps://arxiv.org/abs/2404.11795>MultiPhys: Multi-Person Physics-aware 3D Motion Estimationhttps://www.iri.upc.edu/people/nugrinovic/multiphys/>ProTA: Probabilistic Token Aggregation for Text-Video Retrievalhttps://arxiv.org/abs/2404.12216>BLINK: Multimodal Large Language Models Can See but Not Perceivehttps://arxiv.org/abs/2404.12390>Generating Human Interaction Motions in Scenes with Text Controlhttps://arxiv.org/abs/2404.10685>Dual Modalities of Text: Visual and Textual Generative Pre-traininghttps://arxiv.org/abs/2404.10710>DreamScape: 3D Scene Creation via Gaussian Splatting joint Correlation Modelinghttps://arxiv.org/abs/2404.09227>AesExpert: Towards Multi-modality Foundation Model for Image Aesthetics Perceptionhttps://arxiv.org/abs/2404.0962404/20/2024>Align Your Steps: Optimizing Sampling Schedules in Diffusion Models https://research.nvidia.com/labs/toronto-ai/AlignYourSteps/>Visual Prompting for Generalized Few-shot Segmentation: A Multi-scale Approachhttps://arxiv.org/abs/2404.11732>Partial Large Kernel CNNs for Efficient Super-Resolutionhttps://arxiv.org/abs/2404.11848
>>100140941>>100140837bad wording from my part, that was me paraphrasing the abstractreposting for new threadanons, please help me find this paperparaphrase of abstract: merge checkpoints of different epochs of the same model to reach a better quality than what is possible with gradient descent alone
>>100140958I have a few references to "gradient descent" but nothing to do with model merging
Good evening, anons! I hope everyone is doing well :]
Morning anons>>100141084Heeeeeey
Social commentary
>>100141084Good evening....and good night!
>>100141181Omg heyyyy anon!! :DIt's so great to see you again, as always!!Good morning!! Day just starting for you?Cute Ricedog hahaha! Those eyes!!
>>100140958>merge checkpoints of different epochs of the same model to reach a better quality than what is possible with gradient descent aloneI've been doing this with Loras since forever. Is it news now?
>>100141224Yeah just starting, getting ready to head into work how have you been?
>>100141201Good evening!!! Good night!! Hahahaha!! Great to see you breifly on your way out! Sleep well! Hope to see you tomorrow :]
>>100141084>I hope everyone is doing wellI was until you showed up.
>>100141245You solved a captcha to post this
>>100141236Oh cool!! I hope you have a great day at work! :]I've been doing great! My day just ended like 2 hours ago haha! Well, with work anyways! Tomorrow is my day off so i'm really happy about that!>>100141245I'm sorry to hear that, anon! I hope it gets better soon :]
>>100141278Have fun on your day off :)
>>100141290Thanks, I sure will! :D
>>100141323Really cute witch, anon!I really like that style!
>>100141233it's not a simple averaging, it's something based on the global gradient direction or something
Do you think someday in future I can roam around 70s Tokyo wearing a VR headset? It surely must be possible at some point. There are high detailed maps, satelite images, tons of photography and videos that just need somehow to be geo linked. And for those corners where nothing is exitsting: AI should be able to fill the gaps with likely buildings, etc. Just imagine coming home after work, falling in your sofa, wearing VR and randomly discovering cool places on earth from different times.
>>100141428I don't see why not, anon! I'd imagine you could take old photos and you'd likely get something similar!>wearing VR and randomly discovering cool places on earth from different times.The google earth vr thing is fun! You can just fly around and land somewhere random, it puts you in street mode and you can look around!
100% noob here.How do I improve from here?I think I should inpaint the eyes, but so far all attempts gave me mediocre or outright ugly results.Any tips?
>>100141487When inpainting, you wanna get rid of everything in the proompt besides tags pertaining to the eyes!Otherwise, maybe upscaling? Either way, you're doing well for being new :]
NAI furry v3 has a bunch of bugs right now, some settings are totally ignored>who caresi know some of you are using it.
>>100141472>The google earth vr thing is funThat shit is fire indeed. Probably my number 1 experience in VR
>>100141612I feel the same! HahahaI like finding interesting places and saving them!
Any tips on how to generate realistic selfies?
>>100141681nice to see you back anon, any chance of a catbox?
>>100142001[subject] , selfie or reverse selfie
>>100142001https://civitai.com/models/259627
>>100142241What's the difference?
>>100142327who knows?
>>100142327what model are you using?
>>100142347Juggernaut XL + RunDiffusion 8.0
>>100142327
>>100142549Model? Prompt?
https://youtu.be/eGslweDOihs?si=SvAem-LQXlt5VdHf
>>100142561drunk lady, reverse selfie , dim lighting , after-partyrender, cgi, transvestite , nude nipples , child , skinny , airbrush paseersNeverlandSDXL_v10
Haven't been here in awhile. Is there any consensus on the best XL model? I've tried Dreamshaper but prompt adherence seems worse than the 1.5 derivativesAlso any of you nigs tried Krita diffusion? It's great
spheres, the 2024 fashion statement
>>100142565So far this is the best I can do
>>100142546Awoo
Bunny
>>100141428I’d rather remove all the influences that led to us wanting to return to the world the generation before us knewYou know who that are
>>100142620Rumbling along with JuggernautXL 9 and upwards
>>100142720chain reaction
>>100142713Understandable, have a nice day
>>100142729Looks a little out of place but is a kids playground, so probably standing next to a swing. Been messing around with controlNet and figuring out if img to image is better for certain things if the original image poses are not iconic enough for recognition.
Can you guys create a LoRA for me?
>>100142562>>100142696AWOOOOOO!!!
HyperSD gives pretty good results with autismmix.
>>100141534Okay, tried the inpainting thingy.All criticism, feedback and tips are welcome.
>>100142924Both have their merits, my perception on the meaning between the two is that the purple eyeshadow says "I'm going for a style, a character" and without "look at my huge mammaries"
>>100142824No
>>100140236nice gen op
>>100143022
>>100142824look up for tutorial and make them yourself retard
G'mornin Anons, have a great day!
>>100143078u2
>>100143089
>>100142924This looks great, anon! Good job :DThe eyes are focused in this one!I really like the purple eye shadow! To make it fit better I would make em more smoky than flat! The dress has lots of detail and I think adding more to the eyes would be really nice!It's great though!>>100143078Good morning, anon! You as well :]
>>100143115
>>100143078How do you get faces to actually be like the person?Whenever I do img-2-img it just turns into a random person
>>100143142a lot is dictated by the model but also by denoising str
>>100143142use ipadapter! It works pretty well :]
my heart started pounding when she showed up
>>100143190Nice arcane style! with Lora?
>>100143142InstantID https://github.com/cubiq/ComfyUI_InstantIDUse face
>>100143210yea, I got it dialed down to 0.4 but it comes through pretty strong
>>100143295It looks good imo
considering jumping to comfyUI, i hear there's more flexibility and possibly even a better performance boost than SD forge by this pointwas curious why it stopped getting updated; god dammit.
>>100143205Hers was as wellDue to the trans-fats around her organs and plaque in her veins
>>100143049almost Pink Floyd 90's album cover worthy, nice.
>>100143392>almost Pink Floyd 90's album cover worthy, nice.thanks!
>>100143449Oh this one looks great
>>100143533
>>100143533>Oh this one looks greatPretty easy at https://huggingface.co/spaces/PixArt-alpha/PixArt-Sigma but they throttle you. Anybody figures out how to localize this, let us know!
>>100143628
>>100143672
>>100143686
>>100143635Supposedly you can use PixArt Alpha and Sigma with this extension in ComfyUI:https://github.com/city96/ComfyUI_ExtraModelsBut I haven't tried it myself yet.
>training lora>model looking good>loss slowly dropping>1000 steps into>Loss begins to immediate spiral out of control until only nonsense and nans are left.What the fuck?
>>100143726You can and it's pretty good>>100143635Clone and follow instructions https://github.com/PixArt-alpha/PixArt-sigmaOr use diffusers versionLocal is fully available to anyone actually attempting to use it. Learn to be okay with stumbling.
>>100142304
>>100143635>>100143848and if you don't have 24Gb of VRAM you can use this script to load T5XXL in RAM, but you need at least 20G of RAMhttps://github.com/lrzjason/PixArt-sigma/blob/master/scripts/inference_t5_cpu.py
>>100143892
>>100143848>Clone and follow instructionsdamn conda system hijacking mess.
>>100143953Shit. Happens every time.
>>100143897That's a really nice scene, anon! I like the style too! :]
>>100143722
>>100144042
>>100143726was trying this last night and just got black gen outputs with ExtraModels, at it again this morning
>>100143955It also works with pyvenv or uv. uv is a rust-based pip replacement and stupid fast. Just have to install torch, torchvision, and misc from requirements.txt
My coworker dressed like this today and I couldn't stop staring
>>100144227These look better than the momojiDid you change anything?
>>100144419The character, of course and also I took away some "cute" and made the backgrounds less abstract.Personally, I like the style of the Momiji ones better, I wonder if that's just preference or if my taste is weird.
>>100144615cute fox
>>100144431>>100144615Good morning, Comfy! Great to see you :]]
Good technique for footfags - better alternative to clicking generate over and over and over again until it finally makes passable looking feet:>find (or make with blender) a picture of feet/a foot with a nice shape that you like>outline it with an image editor then place the outline where you want it to be on an image that's the same resolution as the image you're going to generate>make a canny image out of that image>enter the prompt you want while using canny at control weight 0.4, control mode balanced>it will generate the barefoot girl of your choice in the pose you prompted with feet exactly where you want them to be and with the exact shape you want them to beI can't believe I never thought of this before but this works like magic.Now I can gen cute anime girls with feet that look like my favorite foot models.
>>100144701Good morning. Nice to see you too.>>100144668Thank you.
>>100144737What's a canny image?
>>100144865Have you used canny before?
>>100144901I see, so it's the output of an edge detector, right?That's good to know, I don't need to generate feet, but it'll still come in handy. Thanks, anon.
>>100144937Yeah, I'm sure the technique can be applied to other things, like hands.
>>100144898=.= slow at genning as well? Im at almost 8min an image
Just spent 30 minutes trying to figure out what was wrong with 'ModuleList' object has no attribute '1' for the ksampler&adv from a tutorial, was using a sdxl depth controlNet with a 1.5 model as i'd not changed it from the default model when following the yt tutorial. I need a 3rd screen so i dont have to flip back and forth, reee.
>>100144779Thanks! :DI'm gonna try to catch a quick nap hahahaI hope to see you later on!!
>>100145053Take it easy dude!
>>100142667>it isn't going to polish itself
>>100143897>>100143686I like these
>>100145067Thanks! your Watercolor looks good 2
>>100145060I dunno who you are, but I appreciate you, anon! <3 I always see your posts hahaha!Good night!
>>100142825 >>100142546Somehow the colors seem off, still quite awoo!>>100143635Yes, this has woked for a while using the extra models extension in comfyui. Did you get it working meanwhile?
>>100144970>>100145017Why sad? No awoo no more?
you report me because you are a weak limp wristed onionsI post ban worthy images because I am strong, I predate blue boardswe are not the same
>>100145411tiny dick energy
>>100145207the detail is nuts>>100145361good pixel art
>>100145430>zoomer hands typed this
>>100145439I'm a millenial, not a zoomer
>>100145361They are meant to imitate traditional media... is that what you mean when you say the colors are off, or is there more?
>>100144818nauseatingly palatable
>>100145411>I post ban worthy images because I am super stupid, profoundly mentally retardedbeyond all comprehension.Indeed, I believe it. We can see it. Stupidity is your strength and your element. We get plenty of your kind everywhere in the world.And no one wants you all around, shitting up our world with your extreme and worthless stupidity.
>>100145703>post disgusting worthless Warhammer 40k Oh god, leave me alone
>>100145615How should I interpret that?
How do I keep model in RAM?Before each generation, it loads the model in RAM and then unloads half of it after starting generation (not after finishing) and after finishing it unloads the rest. So it goes from 9GB RAM to 15 and then to 12 and back to 9 when finished. It takes 1 min to load it before each generation.Here is info that may be useful.SDXL, Forge, Windows, 4gb vram, 16gb ram.
your circle gets smaller and smaller while mine gets bigger and bigger
>>100145830I can just increase (circle:1.3) weight if I want it bigger. its not magic
>>100145361Soulful
>>100145830Stop buying larger toys pervert>>100145801Would swat last of the hive
>>100145366a-...awoo...
>>100144253Tried this t5 encoderhttps://huggingface.co/PixArt-alpha/pixart_sigma_sdxlvae_T5_diffusers/tree/main/text_encoderTried this t5 encoder based on the ExtraModels suggestionhttps://huggingface.co/DeepFloyd/t5-v1_1-xxl/tree/mainStill black gens. Doing something dumb, I guess, will keep trying.
So supposedly an entire team of professionals tried to make SD try to generate this exact ghost, even with LoRAs, but couldn't do it because AI kept mistaking it for generic ghosts and Pacman ghosts. You guys want to have a try at it and confirm? and shit all over a team of pros in one afternoon? Cause to me it sounds like they're either THAT incompetent or they lied that they tried.
>>100146124nice, i like the traditional media and expression
>>100146156>pls do my work for mehow much are you paying?
"running"
>>100146156How are "we" meant to do that without knowing what it is? Not going to try anyway
>>100144898so cute
>>100146156lol, only if you agree to pay us
>>100146189Thanks. Adetailer left a square for some reason. Not noticed it doing that before.
>>100146156is this how you manipulate your mom into bringing you a bucket of tendies every night?
>>100146156How could it be a ghost if a lora was used? That makes no sense to me
>>100145727that these generated women are so physically appetizing it's hard to look away, but hurts to look, not unlike the feeling you get with images of hyper-palatable foodit's hard to get across the exact feelingthey are 'redpilling', in the sense that they force you awake, into full awareness of the masturbatory and artificial (or maybe more, 'directed'?) nature of these images, as opposed to the illusion of incidentality that most even totally obvious 'lust provoking images' posses
>>100146217tendies are delivered by plate, not my bucket. buckets are for bone-in fried chicken
How do you use wildcards in comfy? I've been using ImpactPack's Wildcard processor, but the node breaks after a while. It stops randomizing the wildcards.Are there any alternatives or fixes?
>>100146235Woah :) well put
>>100146278
>>100145433>good pixel artsigma does it pretty well by default, OTOH it doesn't know many characters
>>100146156>entire team of "professionals" cant into imggen am i supposed to be shocked and surprised
>>100145460Seems a bit off with goal too; perhaps it's a difference in display.>>100145933yea. maybe huge panoramas not too far in the future?
>>100146330neither. its made up because he wants to bait someone into doing his project for him
>>100146338Did she drop her tail?
>>100146156this makes no sensebut also sure as fuck I'm not going to prepare training data and do it; it's work and not within my interests, prepare the training data and put up a request on civitai or w/e if you want a chance at getting it cheap>>100146361yes. common issue on pixart sigma.
Do you think small SD3 models will be popular?
me and my gf
>>100146361And don't ask me why the model learned this.
>>100146416Why "small" SD3 models? Knowing most AI, people will prefer the better performing models even if they take 10 or even 40GB of storage.I think home only users become more bothered at 50GB+ right now and even that might shift in not too much time as people just might buy suitable hardware.
>>100146450coolIt has the fuzz of a real image
So somebody else doesn't waste their time. A1111 api bros. An incorrect sampling method will now return a 404 instead of something useful. With the implementation of schedule type DPM++2M is no longer valid as a sampling method.
>>100146416depends on how it performs compared to SDXL. the small models only purpose is going to be for people who can't run the big ones, but can run more than 1.5. thats the gap SDXL currently sits in, so SD3 has to meaningfully exceed SDXL to make a case for abandoning the existing XL ecosystem while also not getting the full prowess of SD3
>>100146462>Why "small" SD3 models?because 16gb/20gb vram is a big ask and lots of people don't have the hardware for it. a ton of people are in the 8-12 range and will need to use the smaller models
>>100146156
>>100146416would you need a car airbag with that rack
>>100146462exceptional imagery. reminds me of my project to make a new figure lora
>>100146462I don't why somebody may be bothered by storage space.I thought small models are mostly for old GPU users, not to save space on a drive.
Thread challengePost a image of your bakes mid gen, no editing
>>100146563that's not a challenge, are you retarded?
thread challenges need to be more than "hey look at my picture I just made"
>>100146583xhe is, yesI thought that was blatantly obvious tbhonest
>>100145433>>100146463i try to prompt more noise because it adds to the aesthetics (a matter of personal opinion for some themes), but sometimes it ignores me completely.>>100146305the background reminds me of klimt
>>100146583It is, helps anons see how much gets edited for some anons, but if you're afraid I gt it. Lots of anons use detailer passes/inpainting to fix their work before posting
>>100146416Of course if they could compose perfectly well in some ways, it'd be something else.If you could have figure models with a lot of basic indoors/outdoors backgrounds and then add the outfits as LoRa or second model, it'd surely be popular even without the capability to prompt underwater sealife and car parts.>>100146507Everything <=24GB (current consumer hardware) is probably just solved in hardware? People can also use system RAM.YMMV if it is somehow easy to, say, quantize into a small model - then that will of course also get done for about everything. If it however needs a complete separate training and it remains as computationally intensive as checkpoint training is now, I'm not so sure as many will bother to train for much inferior results to achieve <=16GB VRAM, they put in their time and hardware after all.
>>100146061Got it working, needed coffee
>>100146563 I hit skip at about halfway through on the left one.
>>100146601Oh
>>100146615you are mentally challenged challenge
New Thread>>100146577>>100146577>>100146577
>>100146601It's sd's take on Charles Angrand colliding with an anime lora
>>100146636Nice>>100146647I get it you're afraid, I leave this thread with your soul
>>100146661>I get it you're afraid, I leave this thread with your soulyou say the stupidest shit lmao
>>100146563Neat idea. I wish I had gens that didn't have unsafe /g/ stuff that required inpainting. I'll give it a shot if I get a good roll.
Is there anything to discuss in this thread? /lmg/ has good AI discussions, but /sdg/ is just (bad) softcore porn.There's zero development or exchange here. Kinda pathetic desu
>>100146913>lmgall I see there is shitposting about people I don't care about. Seems the generals match to me.
>>100146913why do you say this retarded shit? comfy, ani,city86 that's more dev in one fucking pinky toenail than those other shitty places
>>100146913What is there even to discuss at this point?Ultimately, all the interesting development happens in the development of new base models - actually new ones that aren't just "Pony, but mixed with the other models, but at slightly different percentages than all the other mixes!"Other things, workflows, plugins etc. are interesting, but rarely useful. Usually it just increases the amount of misgens by 200% in exchange for making it 15% easier to prompt for something specific. Yawn.
>>100147014>>100146913>>100147131You can always try creating an Aavegotchi and a Lickquidator correctly.
>>100140236dew+, (tr)ani (unemployed, won't ever be japanese or in japan) and (cum)fy are niggers, worth less than the control net chink