Previous /sdg/ thread : >>101412216>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 news07/15/2024>Google's Gemini AI scans Google Drive PDFs without permissionhttps://www.tomshardware.com/tech-industry/artificial-intelligence/gemini-ai-caught-scanning-google-drive-hosted-pdf-files-without-permission-user-complains-feature-cant-be-disabled>ComfyUI-ppm: Fixed AttentionCouple/NegPip, more CFG++ samplers, etchttps://github.com/pamparamm/ComfyUI-ppm07/14/2024>ControlNet-Union-SDXL updated, adding Inpaint and Tilehttps://github.com/xinsir6/ControlNetPlus/tree/main>ComfyBlog July: Versioned Releases, Beta UI, New Modelshttps://blog.comfy.org/new-models-versioned-release-beta-ui-new-models/>danbooru2023-florence2-captionhttps://huggingface.co/datasets/KBlueLeaf/danbooru2023-florence2-caption>Nvidia RTX 50 graphics card family TDPs 'leaked' by Seasonichttps://www.tomshardware.com/pc-components/gpus/nvidia-rtx-50-graphics-card-family-tdps-leaked-by-seasonic>GLSL Nodes for ComfyUIhttps://github.com/patriciogonzalezvivo/comfyui_glslnodes07/13/2024>Kolors + Controlnethttps://github.com/MinusZoneAI/ComfyUI-Kolors-MZ>OpenAI blocked staff from airing security concernshttps://thehill.com/policy/technology/4770116-openai-blocked-staff-from-airing-security-concerns-whistleblowers/>ToonCrafter_with_SketchGuidancehttps://github.com/mattyamonaca/ToonCrafter_with_SketchGuidance>DiffSynth-Studio Adds support for training Kolors LoRAhttps://github.com/modelscope/DiffSynth-Studio/blob/main/examples/train/kolors/train_kolors_lora.py07/12/2024>Progressive Growing for Stable Diffusion Web UIhttps://github.com/rmcc3/experiments-progressive-growing>Video Diffusion Alignment via Reward Gradienthttps://vader-vid.github.io/>MAVIS: Mathematical Visual Instruction Tuninghttps://github.com/ZrrSkywalker/MAVIS>Pallaidium Adds Kwai/Kolors Supporthttps://github.com/tin2tin/Pallaidium#change-log>SEED-Story: Multimodal Long Story Generation with Large Language Model https://github.com/TencentARC/SEED-Story
>mfw Research news07/15/2024>TCAN: Animating Human Images with Temporally Consistent Pose Guidance Using Diffusion Modelshttps://eccv2024tcan.github.io/>LAPT: Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Modelshttps://arxiv.org/abs/2407.08966>Semantic UV mapping to improve texture inpainting for indoor sceneshttps://arxiv.org/abs/2407.09248>A Fair Ranking and New Model for Panoptic Scene Graph Generationhttps://arxiv.org/abs/2407.09216>Evaluating the Adversarial Robustness of Semantic Segmentation: Trying Harder Pays Offhttps://arxiv.org/abs/2407.09150>Refusing Safe Prompts for Multi-modal Large Language Modelshttps://arxiv.org/abs/2407.09050>Full-Stage Pseudo Label Quality Enhancement for Weakly-supervised Temporal Action Localizationhttps://arxiv.org/abs/2407.08971>Lite-SAM Is Actually What You Need for Segment Everythinghttps://arxiv.org/abs/2407.08965>One-Shot Pose-Driving Face Animation Platformhttps://arxiv.org/abs/2407.08949>AirSketch: Generative Motion to Sketchhttps://arxiv.org/abs/2407.0890607/14/2024>Feature Diversification and Adaptation for Federated Domain Generalizationhttps://arxiv.org/abs/2407.08245>Synchronous Diffusion for Unsupervised Smooth Non-Rigid 3D Shape Matchinghttps://arxiv.org/abs/2407.08244>SoupLM: Model Integration in Large Language and Multi-Modal Modelshttps://arxiv.org/abs/2407.08196>Foundation Model Engineering: Engineering Foundation Models Just as Engineering Softwarehttps://arxiv.org/abs/2407.08176>Multi-Object Hallucination in VLMshttps://arxiv.org/abs/2407.06192>Vision-Language Models under Cultural and Inclusive Considerationshttps://arxiv.org/abs/2407.06177>MobilePortrait: Real-Time One-Shot Neural Head Avatars on Mobile Deviceshttps://arxiv.org/abs/2407.05712>VideoCoT: A Video Chain-of-Thought Dataset with Active Annotation Toolhttps://arxiv.org/abs/2407.05355>CountGD: Multi-Modal Open-World Countinghttps://arxiv.org/abs/2407.04619
what's the go-to method for describing subjects separately these days? I'm trying to make hentai sex images
>>101421090they fuckin
>>101421081newfag
>>101421101
>>101421066seperation the the prompt .. describe one subject, decribe the scene, at persons prompt like 2girls, describe second subject .. mostly works for me, sometimes the subjects get garbled together, but mostly it works .. also if you wanna make porn you should learn booru tags, if you can stand watching at unspeakable porn for a while by looking what tags are used on danbooru or gelbooru for multiple person stuff
20 days finished plus half the next 10over 9 million imagesvery nice
testing out make a penny arcade lora
>>101421105oh?
>>101421142sexists??
>>101421119sorry I should have been a bit more clear. I do know all the basics of how to make a scene involving two people. I'm wondering what the current meta is for saying "he's wearing a red shirt, she has blonde hair" etc. I know there's some prompt masking method but it never worked that well for me. There's also AND, but, again, mixed results. Prompt area limiting produced very bad jank. Etc. I don't do this kind of fine-grained control most of the time so I don't know if the methods have improved.
>>101420852>>101419843>>101419784>>101419369
>>101421164aslong we are stuck with CLIP as text interpreter I don't think there ever will be a really definitive way of solving the problem other than regional prompter (which doesn't work on SDXL afaik?)eventually txxl5 will solve that .. I tested it earlier and it is smart enough to that you can give characters names or numbers then tell it "person one has blue hair, green eyes. person two is blonde and is wearing a trenchcoat." .. but the problem with txxl5 is the ram requirements and that the SD3 checkpoint just sucks..also AND will just mix the left side of the prompt with the right side, so its the exact opposite of separation
>>101421157yeah it doesnt need exists in this case, usually its exists false because the field to mark something as complete/searched or whatever is added after its done and i query for documents where it doesnt exist. just the way ive ended up doing it, otherwise i'd need to modify the documents to add the fields before insertion and i might not know what extra fields i want until im working on the next stage<3 mongodb desu
>>101421209https://suno.com/song/58d304b0-6061-4b20-a0d3-e2dfa202fb41
>>101421408funny
>>101420987cute rat, i saved it
>>101421209>also AND will just mix the left side of the prompt with the right sideNot exactly. But in effect that's usually what you get. Using AND correctly is hard, which is why I don't.
>>101421494
>>101420987>>Discord >6wUwtcJsr2Wat is?
>>101414994Base was Kolors then used SD3 with Ultimate SD Upscale.
Sorry in advance because this has been asked 1000 times already, but I've been thinking about training a LORA on some of my old pictures. I assume this would be a style LORA. Is there a recommended settings and workflow for this? Can I just auto-caption them with like one line from CLIP and my trigger word, or is there a better way? It's probably like 100 pictures of people, and I was thinking 1.5 (more specifically, realisticvision40 or whatever the default is with forge) under the assumption that it's quicker. Not worried about interoperability because i'm not sharing it
>>101421600do not under any circumstances click that linkunless you're retarded—good riddance then!
>>101421859yes, I'm not sharing it with you
>>101421977NO SPOONFEEDING
>>101421974r the good posters there?
I'm never joining a 4chan groomercord, it's either going to get me on a list or contain stuff that should put you on a list
>>101421930noticed
>>101421980Cool, which artist lora(s) did you use for this one?
>>101422039I am working on a penny arcade lora as a bait project, it isn't very good yet as I did the lora with like 20 panels so it's mostly just autismmix source_cartoon doing the work
You're already on a list faggot
>>101422082who are you? Nixon?
>>101421930dude.. blue board, no female nipples allowed, the Americans get their panties in troubles if you post em
>>101422128
>>101422378ya ya.. now go wash em, you soiled em
damnit.. leaked tdp values dont bode well.. 500W 5090 ..
>>101422488nice style
>>101422517thank you
>>101421859https://youtu.be/7m522D01mh0?t=657>assumption that it's quickerThat what is?
>>101422516I used two, 290x2 R9 8gb's for awhile, each one was like 600wat and would kick 15amp breakers. Had to put a plugin off of the dryer circuit.Knowing NV they'll do something insanely shitty like limit the 5090 to 12gb vram or something.
>>101422578rumors say 28GB for 5090 .. which would really suck, cause 2000€$ for an upgrade of 4GB from from my 4090 with 24GB would be really hard to argue, even if I get the 4090 sold reasonably ..
>>101422620Yeah that'd be a hard sale even if one had the money to just splurge.
>>101422488She tried her best.
>>101422737rofl
>>101422737>it's another stepmother stuck in the dryer video
https://colab.research.google.com/github/hollowstrawberry/kohya-colab/blob/main/Lora_Trainer.ipynb#scrollTo=OglZzI_ujZq-in case some1 is looking to make a lora, this colab seems to be working, and simple to use
cant sleep desu
>>101422988sleep! or she comes to visit you
>>101422988Let's see your best
Well, I guess you can have okish hands at the cost of WTF.
>>101423039HORSEPUSSY
>>101423005it would be nice to have someone visit desu>>101423019this is a scrape of playground, i dont really gen myself, i like this one i made testing attention scale in diffusers, it allows below base resolution on xl which is pretty neat
>>101423039she agrees>>101423048no its a midtaur
You ever find an image that you like so much you spend hours trying to re-create it perfectly, but it's an impossible goal to reach because it was probably made using custom loras you'll never have access to?
This was meant to be a "monster girl". Definitely ended up waaay more "monster" than "girl" and I don't even know how>a girl at the beach, digitigrade legs, monster hands
>>101421148>penny arcade lora>doesn't look like the artist had a strokeGenuinely makes for a great style
>>101421148I made a dofantasy lora, but it was just weird
>>101423279reminds me that the cwc(sonichu guy/gal) lora has a nice effect/quality, without the full horribleness of the actual art
What do I have to do to get comfyui speeds with my own python script? I tried out some optimizations but it still takes 3 times as long.
>>101423317do what zhang did, Ctrl c, Ctrl v, then call it your own code
https://litter.catbox.moe/mhtnc5.pngI don't even know how this happened. Probably because of the typo in the prompt
What's the free best ai image editor to quickly change parts of the image?I have this photo and I want to tell the ai to change the outfit from this slutty mess to a proper school uniform.
>>101423435Looking good
>>101423118I even do that with my own shit.>"Oh god she's perfect.">48hr later and I'm binge drinking while staring at the GUI in defeat
>>101423460ty
>>101423410probably krita and the plugin since everything else is webshit
>>101423118>>101423470Same hear. Accidentally gen something then spend way too long trying to build off of it
Why do so few people, if any, inpaint bad details to make it look better?
I kinda like how you get the impression that most of these digitigrade ladies look really freaking tall despite nothing to compare them to
>>101423547in this space attention is all you need. the problem is, we are all attention deficit
>>101423570wouldn't you get more attention if it looked better?
Is there any animation anon here?I want to use animatediff + controlNET on comfy but i am not sure where/how to start.The rentry on OP were last updated like a year ago, before most of the animation improvements for local AI.
made a new lora
>>101423675>LongloliXL
Christmas in July
>>101423681bleh, it is an sd 1.5 lora, actual a locon, my bad
Tried getting digitigrade legs but with normal human toes. Turns out the result isn't that much different from plain-old tiptoes. Disappointing
>>101423692Looks good though.
mite b kewl if there were a way to control size of objects, like (Meatball;SIZE:25%)I.e 25% of the image's dimensions are meatball.
Mistakes were made.
>>101423800there are no mistakes. only happy accidents
>>101423593what are you trying to do? rotoscope or interpolation?
>>101423895imagine the bug bites
>>101423497Ok I dl Krita and tried this with local model using 4060.Pros:Easy to install, few clicks and about 10 gb of disk space to run the model locallyDefault anime model is decentEasy to select portion of the picture to inpaintCons:Disk space usageSlow to generate without high end gpuis Krita generally better than paint.net for general image editing?
>>101424017Paint.Net with all the extensions and shit is pretty great.CS6 works under Wine now perfectly though.
>>101424017>4060its so ogre
>>101424055Why?
>>101424042I have used paint.net with a lot of extensions like TR's alphacutter and DistortThis! for making most memesBut quick recolors are hard and the layer system is too simple
>>1014240854060 is pretty wimpy for animation stuff, esp since you wanna have controlnet in the mix. thats a lot of vram you don't have
>>101424101I haven't dl a controlnet, what do those do?16gb vram (ti)
>Imagine a young girl lying on the grass in a park. Her head is framed by a speech bubble coming from her mouth that reads "I can't make this girl lie on grass." The girl is wearing an outfit with a sign written on it that reads "SDG" (Sustainable Development Goals). She has long, black hair and is smiling in a psychotic way. The background is monochrome, with high contrast and highly detailed shading. The image is taken from a 2000s-style comic book with artline duotone drawings. It's shot from a 3/4 angle at a portrait perspective to show the girl's face in detail. Just realized each model has different min denoise for HiRes fix upscaling. Like this model still unstable with 0.56
>>101424110oh, I thought you were this guy >>101423593>I want to use animatediff + controlNETalso the ti was an important bit you left out lol. thats not a bad card
>>101424258cool scene. you know she's a good hacker because of how many monitors she has
love tuna casserole
>>101424353Honestly? I've never met a casserole I wouldn't eat.
>>101424356true enough.. wait, nevermind, i have never tried the green bean variety, it always repulsed me by sight
>>101423593Get this:https://toyxyz.gumroad.com/l/ciojzYou can animate models in Blender and get controlnet image sequences out of them. You can use those withhttps://github.com/Kosinkadink/ComfyUI-Advanced-ControlNetnodes. And for animatediff, use https://github.com/Kosinkadink/ComfyUI-AnimateDiff-EvolvedAlso check out Mixamo for free animations that you can retarget to any humanoid model. It's cloud shit but it's free if you get an Adobe account.
>>101424384>>101423593Or you can run pose detection on existing videos of humans (or of yourself, but the body ratios won't match), but it's a bit finicky.
>>101421489lora used for this gen? i have issues getting eyes like that
any nice pinups?
Neeeeeeeeeeeerrrrrrrrrrrrrddddddddddddsssssssssssssssss
>>101424384if you're already making fluid animations in blender, why not just make the whole scene in blender and v2v it? then you have lighting control too>>101424410no ulong time no see nerds guy
>>101424415Been busy unfortunately
pinup, but maybe not nice
244 pre-prints@_@mondays are crazy
Good evening, anons! I hope everyone is doing well :]
>>101424780Heya Debo! Great to see you as always!Omg that's such a cool style! :DI saw your other gens above mine and they look awesome!>>101424545The words on this one came out so good! LOL Debo Dread hahaha!
>>101424800this PW series is supposed to have "PW World" but it doesn't like words as much even though its mostly the same prompt
>>101424807Omg I love this one!!! It looks so cool!The words are super close too!!That outfit looks amazing!
is civitai slowly dying? few months ago they made that generating uses their 'buzz' currency thing. Few weeks ago they halved the amount you get per day to 25. Few days ago there was an 'update to generation cost calculation' and I noticed that for the last thing I made price increased from 8 to 10 per batch. So you get 2.5 batch a day, or 5 if you keep publishing a picture every day. That's pretty much nothingluckily I do not need to use it, been genning locally and just clicking the get buzz button when I remember to 'just in case', but can the service survive if it becomes pretty much fucking useless? Csn it attract enough paypigs?memes aside it is pretty handy as a lora repo and propmpt inspiration
Are RTX 3090s and 4090s still the gold standard for local models?
>>101424908I told you, this prompt is cookin up some super cool PWs>>101424918>is civitai slowly dying? I think they're losing ground on both the user and and the business end. the buzz thing started with their pivot into trying to become profitable and monetize the platform. they get skimpier and skimpier as they aim to squeeze out more revenue and make their numbers add up (not happening). at the same time, the SD3 catastrophe killed a stream of content as well as general user engagement. combined with people getting bored of imggen, they've prob seen a huge dropoff of both submissions and page visits
>>101424966I get much better generation performance with my 970
>>101424978The site is slow, awkward and laggy in general, huggingface has 90% of anything relevant and you only need so many models. (Though I'll confess I'm at like 150gb now)
>>101424978That's so cool! I love the expression! It really is hahaha>>101424966Pretty much haha it's super fast! :]Also lets you make bigger pictures!
would using an external gpu over thunderbolt 3 bottleneck me?
>>101425072I would doubt it really. VRAM is all that actually matters. And xformers if you're on NV. (Which you probably are/should be.)
>>101425054>The site is slow, awkward and laggy in generalwell thats not a new thing. its always been a poorly built site. my suspicion has been that they're infrastructure guys who frankensteined together a site>>101425055lots of wildcards too. PW comic
>>101425072for gaming yeah, for genning I dunnoon paper it's 1 string up and 1 picture down, every 30 or so seconds. Doesn't sound like it's using all the bandwidth available
behold, my new and very unique OC: PJ. please donut steal
Replay'an this with New Threat 2.0 mod and it's the best thing ever
>>101425106What model are you using? That's an amazing style!>>101425152LOOOL imma steel :]
>>101425170its a model called boomer art model. not to be confused with coomer art model. VERY DIFFERENT!https://civitai.com/models/163139/boomer-art-model-bamit works really well with an artist wildcard list I modeled off of a bunch of artist names illuminati anon gave me
>>101421113cute meese
>>101425200there should be an animorph loraturn Azula into a wombat
>>101421494um.. um... c-catbox??
>>101422644Holy moly that's great. There a D&D cover art lora somewhere?
>>101425214>there should be an animorph lorahttps://civitai.com/models/123036/animorphs-cover-sdxl
>>101425250lol
>>101425250where my picdurr
>>101425265even the archive didn't want it O_o
>>101425269Don't think I've ever saw that happen before
>>101425243No, it was a Luis Royo Pony lora
>>101425250I can't even re-up it cuz it says its a duplicate. here's the jpg version. I will not be denied
>>101425285cool
>>101425308bad ass
>>101425308moggs me both technically and bone structurally
>>101425200LMAO I would ate to make that mistake hahaha!!!Thanks Debo! I'll give it a shot :]I should probably try using wildcards haha I never really gave that a go
>>101425330>I never really gave that a goreally?? you get so much variation from your gens that I always figured you have some wildcarding in your stuff. I'm surprisedyou actually probably already have a wildcard node installed. if you have the impact node pack, it comes with a wildcard processor with all sorts of different ways to approach wildcardinghttps://github.com/ltdrdata/ComfyUI-extension-tutorials/blob/Main/ComfyUI-Impact-Pack/tutorial/ImpactWildcard.md
>>101425351Yeah hahaha! I do have the impact node pack actually! I don't have it in my workflow tho I don't think thoI should play around with it and see what happens hahaYou said you have a list for em? How would I use something like that?Omg that hair is SO cool!!
>>101425395>How would I use something like that?there's a wildcard folder in the custom node's folder and you just put txt files there. so it'd maybe be something like artists.txt, then the file would be>artist one>artist two>artist threethen in the wildcard processor, you invoke it with __artists__. you'll need to hit 'refresh' anytime you add a new wildcard file. when it reads it, you'll be able to see the value it selected in the 'populated' segment of the node.you can also do inline wildcarding like how a1111 supports, like {artist one|artist two}
Bliss.
>opens thread>no Aryan 1girls>closes thread
>>101425455>Bliss.I understood that reference
>>101425460
pw from an alternate realitythe silly hat reality
>>101425395Ohhh cool! I'll try that out!!I noticed with Kolors if you add '|' it does something hella weird and creates an error haha>>101425512I love that silly hat haha it's like a really cool beret! Looks like she's talking to someone on an earpiece too!
>>101425551bizarro-pw is all about the bizarro head gear
>>101425570what color would she be in magic the gathering? the art style looks like it could belong
>>101425597good eye! I have some MtG artists scattered amongst the wildcards. magic users tend to fall into blue cards but she could also be summoned with black mana if there are any nogens on the field
>>1014256321am = F5, elsewhere
>>101425642I wanna get the news up before I go to sleep but I'm starting to regret it
>>101425657So press F5 then...
>>101425327thank>>101425329lel
>>101425657only 15 images
>>101425659ty king
>>101425694have a good one!
>>101420987Question as a complete retardDo I get everything from beginner UI install and local install? Or some one specific thing and stick with that?
>>101425694catbox before you go?
>>101425717https://files.catbox.moe/7th5bl.png
>>101425726bye
>>101425713you only need to get ONE of the UI from the beginner section. Just try out for now, make sure your rig is up to it and then after a while when you get familiar with things then you can try out the non-beginner ones.
>>101425779nice Harley booba
why are sdxl backgrounds so flat and soulless
>>101424415>v2v itFor one thing, you don't need to have an exact model. If you're using the pose controlnet only, you can grab any model that has roughly the desired body ratios, use the attach version of the toyxyz thing, and get the pose images. If you're using other controlnets too, the model may need to be a bit similar to you end output, depending on what you need (for example, if you're using the lineart or depth controlnet, for the hands for example, it'll need to match, at least for the parts of the body you're using it for).Also, I find that v2v looks too rotoscopey (but I usually try to make anime with it, for realistic it might matter less).
>>101423925you are garbage
>>101424918imagegen saas isn't profitable without a strong paid base. the 0.1% who buy buzz doesn't come close to offsetting the cost of free users. they do training too which has higher costs. buzz is also highly susceptible to bots. they fucked up, basically. imo they should have taken the faux non-profit route, it would have ironically been more profitable
image ids on civit go to 19m or something but there's barely 5m left because they delete everythingthat's 19m lifetime when real saas like playground are doing like 12m a month publicly plus probably about the same private>>101425106i've got the source, it's not too bad, just the usual nextjs mess
has amd gotten any better recently for proompting or is it still shitty?
>>101426481no, maybe single percentage figure gains. definitely do NOT buy an AMD gpu for AI even if you see a really good deal.
>>101426501alright damnGot a 6700xt in my main rig and a gayming laptop with a 3060 but i honestly dont like getting on the laptop Crazy how much the mobile 3060 with half as much vram outperforms my amd card
Last one from me, good night anons
>>101426667gn
>>101426909oh lordsetup?
>>101425638>>101425638>>101425638