Previous /sdg/ thread : >>101047757>SD3 info & downloadhttps://rentry.org/sdg-link#sd3https://education.civitai.com/quickstart-guide-to-stable-diffusion-3https://aitracker.art/viewtopic.php?t=57>Beginner UI local installFooocus: https://github.com/lllyasviel/fooocusEasyDiffusion: https://easydiffusion.github.ioMetastable: https://metastable.studio>Local installAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiComfyUI: https://github.com/comfyanonymous/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 forksSD.Next: https://github.com/vladmandic/automaticAnapnoe UX: https://github.com/anapnoe/stable-diffusion-webui-ux>Run cloud hosted instancehttps://rentry.org/sdg-link#run-cloud-hosted-instance>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://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
Cursed thread
>mfw Resource news0619/2024>TroL: Traversal of Layers for Large Language and Vision Modelshttps://github.com/ByungKwanLee/TroL>diff-sampler: open-source toolbox for fast sampling of diffusion modelshttps://github.com/zju-pi/diff-sampler>Diffusers Adds≈ SD3 ControlNet and Multi-ControlNet Supporthttps://github.com/huggingface/diffusers/pull/8566>Better & Faster Large Language Models via Multi-token Predictionhttps://huggingface.co/facebook/multi-token-prediction06/18/2024>Release of training script of PCM-LoRA with Stable Diffusion 3.https://github.com/G-U-N/Phased-Consistency-Model#news>Autoregressive Model Beats Diffusion: Llama for Scalable Image Generationhttps://peizesun.github.io/llamagen/>Flash Diffusion: FlashSD3https://huggingface.co/jasperai/flash-sd3>SD3-Controlnet-Tile https://huggingface.co/InstantX/SD3-Controlnet-Tile>Sharing new research, models, and datasets from Meta FAIRhttps://ai.meta.com/blog/meta-fair-research-new-releases/>The next chapter for ComfyUIhttps://blog.comfyui.ca/comfyui/update/2024/06/18/Next-Chapter.html>DeepFuze: Deep Learning Tool Integrating Lipsync, Face Swap, VidGen and morehttps://github.com/SamKhoze/ComfyUI-DeepFuze>Glyph-ByT5-v2: A Strong Aesthetic Baseline for Accurate Multilingual Visual Text Rendering https://huggingface.co/GlyphByT5/Glyph-SDXL-v2>MegaScenes: Scene-Level View Synthesis at Scalehttps://megascenes.github.io/>microsoft/Florence-2-large - New 0.23B | 0.77B Modell for image captioninghttps://huggingface.co/microsoft/Florence-2-large06/17/2024>CivitAI: Temporary Stable Diffusion 3 Banhttps://civitai.com/articles/5732>VEGAIcon: Learning Interleaved Image-Text Comprehension in Vision-Language Large Modelshttps://github.com/zhourax/VEGA>ComfyUI-LuminaWrapperhttps://github.com/kijai/ComfyUI-LuminaWrapper>Generating audio for videohttps://deepmind.google/discover/blog/generating-audio-for-video/
>mfw Research news06/19/2024>Synergizing Foundation Models and Federated Learning: A Surveyhttps://arxiv.org/abs/2406.12844>LayerMerge: Neural Network Depth Compression through Layer Pruning and Merginghttps://arxiv.org/abs/2406.12837>VIA: A Spatiotemporal Video Adaptation Framework for Global and Local Video Editinghttps://arxiv.org/abs/2406.12831>Neural Approximate Mirror Maps for Constrained Diffusion Modelshttps://arxiv.org/abs/2406.12816>AITTI: Learning Adaptive Inclusive Token for Text-to-Image Generationhttps://itsmag11.github.io/AITTI/>Extracting Training Data from Unconditional Diffusion Modelshttps://arxiv.org/abs/2406.12752>AGLA: Mitigating Object Hallucinations in Large Vision-Language Models with Assembly of Global and Local Attentionhttps://arxiv.org/abs/2406.12718>SUPER: Selfie Undistortion and Head Pose Editing with Identity Preservationhttps://arxiv.org/abs/2406.12700>Do More Details Always Introduce More Hallucinations in LVLM-based Captioning?https://arxiv.org/abs/2406.12663>Mixing Natural and Synthetic Images for Robust Self-Supervised Representationshttps://arxiv.org/abs/2406.12368>Immiscible Diffusion: Accelerating Diffusion Training with Noise Assignmenthttps://arxiv.org/abs/2406.12303>COT Flow: Learning Optimal-Transport Image Sampling and Editing by Contrastive Pairshttps://arxiv.org/abs/2406.12140>ARTIST: Improving the Generation of Text-rich Images by Disentanglementhttps://arxiv.org/abs/2406.12044>Not All Prompts Are Made Equal: Prompt-based Pruning of Text-to-Image Diffusion Modelshttps://arxiv.org/abs/2406.12042>Decomposed evaluations of geographic disparities in text-to-image modelshttps://arxiv.org/abs/2406.11988>MAC: Benchmark for Multiple Attributes Compositional Zero-Shot Learninghttps://arxiv.org/abs/2406.12757>Disturbing Image Detection Using LMM-Elicited Emotion Embeddingshttps://arxiv.org/abs/2406.12668>What makes two models think alike?https://arxiv.org/abs/2406.12620
funny how debo always claims how its all lies and shit, yet he cannot refute even a single point of the pastebin
>>101052954>crave attention>get negative attention by attacking devs and model makers>beg comfy for scraps of clout>post personal information on 4chan
should I prune my character focused dataset of useless tags like "nose, weapon, duck, pillow, etc.." that wd14 throws in there in large amounts?
>make SD3 tensorrt engine>try running it using workflows from https://github.com/comfyanonymous/ComfyUI_TensorRT/pull/30>picrelwat do? it looks like a vae issue but i saved it from the SD3 Medium checkpoint
>>101054016>the SD3 Medium checkpointlooks like its working?
>>101053968Why would he? The hastening is straight bullying.
>>101054037>hasteningphone post confirmed
>STILL NO LLM INTEGRATION TO AUTO FIND THINGS AND JUST DO WHAT YOU WANTholy shit how hard is it to have this
>>101054037I could argue/debate/refuse every single line of it but the people who would give me an honest shake don't care in the first place and the people who are motivated by hate wouldn't bother giving anything I say an honest read. I've pushed back on many items in the past and they still pretend its gospel. so there's no motivation or reason to bother wasting my time or energy in what would be a completely pointless activity
>>101054098You never tried LMAO
>debo derangement syndrome in action a tale as old as time
>debo still seething over his exile
>i could argue every single point>but i wont
>ran phoneposting >truly organic
>its all one anonWe're all laughing at (you)
should i join the discord?
>(((we)))
>>101054133>doesn't read what I wrote>doesn't engage with what I wrote>ignores past events entirely>casts blame on meexactly my point. why would I waste time on you? you're not an honest actor and have zero intention ever to be. last (you) for you btw
Should debo make a counter-pastebin? y/n
>i could argue every single point>but i wont because [insert made up reason here]
>lives rent free
>>101054016make sure to use the batch size the engine is intended for
well I just blew 2,500 on a hail mary attempt at training a realistic dataset on zonkeyrealism and it was a complete failure, utterly unusable. I need a a new gpu so bad, the 50 series needs to hurry up and release.
>>101054367what are the odds that your next attempt would be a success?i have 34,000 i'm not doing anything with.
>decide to check out sdg again>thread's filled with some schizobabble about debo againAre those bots?
>>101053932>>101054017Cool
Trying to make a wizard rap album. SD3 is almost as bad at generating guns as it is at generating women.
>>101054222He won't
>>101054367can you show us some of the abominations it produced?
>everyone that hates me is ran
>>101054432He fucking will
>>101054489And I'm actually a woman
Silly beach
maybe make it have a pimp staff (cane) rather than a gun give it's a wizard
i am schizo anonask me anything
>>101054474jokes on him, everyone hates ranfaggot too
>>101054436>>101054336>>101054222>>101054063nice
>>101054307i just regenerated the engine using picrel
>>101054541Hi debo
>>101054307>>101054562still getting this
>>101054575show workflow of the image gen
>open thread>still the same debo shitgoodbye forever, hope yall rot in this shithole
>>101054367oh no, did you at least learn or get something out of the experience?
>>101054624>goodbye foreverkek good one
>>101054624ok, see you tomorrow
>>101054624>spam debo posts>omg debo posts again what a shit thread>im going to leave sdg forever>spam debo posts>omg debo posts again what a shit thread>im going to leave sdg forever>spam debo posts>omg debo posts again what a shit thread>im going to leave sdg foreverhm
>Debo is crying
>>101054117>reposting someone else's workThat's guro btw.
>>101054595
>>101054700i'd test with a prompt.also what happens if you load the SD3clip or SD3T5XXL checkpoint and use its CLIP and VAE instead?
It's been a fucking YEAR of debo obsession.It's actually fucking insane man.
>>101054745>i'd test with a prompt.still not working>also what happens if you load the SD3clip or SD3T5XXL checkpoint and use its CLIP and VAE instead?dont have it right now, will download, thank you for trying to help
this checkpoint merged the L clip and g clip from SD3 into its SDXL model? does it really make a difference? https://civitai.com/models/469255/pixel-alchemy
>>101054777crazy how he doesn't realize no one likes him once they get to know him, right?
>>101054809I don't even care that he's a slightly annoying goober.I mean... is it really worth your time?
>>101054809He got exiled from the discord
>>101054830>slightly
>>101054866>mfw no one waters me
>Trani and PW>Troons>only faggots that like debo
Beach day
>>101054981Wait didn't trani kill himself?
>>101054981He's a useful idiot I'll give him that. Probably the only reason Trani keeps him under her wing.
>>101054745>>101054783
I want to recreate classical paintingswhat is a good chekpoint/lora for this?
>>101055029Many SDXL checkpoints are very influenced by anime and sameface. If you want to avoid that use Dreamshaper XL turbo, it's better balanced. Then there's Rayna XL which is trained in art. The turbo version is paywalled, but the regular XL isn't.
>>101055098>dreamshaper
>>101055029I didn't see the Lora question. I use the John William Waterhouse Lora all the time. But my advice is to use 5-6 artists tokens at a time and play with weights and order to find a sweet spot.
Florence 2 is pretty okay in my experience so far, it starts the tags with "the image is of" , but you can grep that outFor the size and vram requirements it's beating the shit out of anything I've used so far for nlp processing
>>101055029Did you actually know SDXL knows thousands of artists you dumb fuck? You could get away with vanilla but I guess you are too much of a promptlet for that.
>>101055029what artist?
>>101054415I really don't know. my dataset is fine. I just wanted to try training a realistic lora on pony and I can't trial and error test when i'm spending credits. I probably should have trained it on the more vanilla ponyrealism model, or even base pony.I'll probably go back to making more simple concepts that only cost like 600 credits and shelf this for now. are you paying for credits? i've just been claiming the free ones everyday
>>101055273Did you know if god tried with all his might he couldn't make a bigger faggot than you
>>101055301David Cherry, of Age of Mythology fame
>>101055301sdxl artists:https://proximacentaurib.notion.site/e28a4f8d97724f14a784a538b8589e7d?v=42948fd8f45c4d47a0edfc4b78937474
>>101055273Masterpiece1girl, big boobart by Alphonse Muchaart by Greg Rutkowkskiart by Leonid Afremovart by artgermtrending on artstationhmm, let me see8k, high definitionoh yesi am the prompteurre
>>101055375thanks
>>101055322nah i've just been a supporter since february 2023.it's a website i regularly download shit from and use constantly, so i support em.same with 4chan desu, i got the pass too.
>>101055322the only thing i ever spend it on is tipping random people whose models i like and use often
>>101055400yes
>>101055610real nice6 fingers tho
>>101055333Nice succubus
>>101055624i really want to slap dat ass
>>101055640>>101055663Thanks
>>101055681>>101055610what are you using?
>>101055610Fixed
>>101055720cyberrealisticXL_v20but I bet most XL models will do similar>>101055731cool
>debo nogen grooming newfags
>>101055400I call myself an ateur. Prompteur is so superficial anyway.
>>101055098>Dreamshaper XL turboany suggestion on the VAE?
>>101055921
>>101055979Annuit cœptis
>>101056025brown eyes bro
What's a good all-around model for doing all different kinds of generations? I try out different models, but I always keep going back to crystalclearXL
is there such a thing like a model interrogator? paste an image, it runs an interrogator and tries to line it up, taking into account the checkpoint you're using?
physical attribute hierarchy before 2 years of SD gooning:>tits (big)>ass (big)>face >waist>thighs (fat)>legsafter 2 years of SD gooning:>hips>fat distribution/tone>face>shoulders>waist>neck>legs>tits (small to medium and well-shaped)>butt (slender and proportional)someday hopefully SD refines further so that I can get deep into hands and feet
>>101056220pretty good anon promt?
>>101056468https://files.catbox.moe/l55cui.png
Challenge!https://desuarchive.org/g/search/subject/%2Fsdg%2F/type/op/start/2022-08-01/order/asc/Browse an early /sdg/ thread, find a nice or antinice gen
>>101056436physical attribute hierarchy>tummy>everything else
I heard that the reason SD3 still sucks is that Stability has gone full woke, they had mangled and disfigured people in the training data and REFUSED to cleanse the data or tag "maimed hands, disfigured face"/etc. because it would "perpetuate ABLEISM". Much like Dall-E FORCES you to look at "ethnically ambiguous" people SD3 forces you to see hideously deformed creatures because "ETHICS"
>>101056366Gonzales WetDreams? AutismPony?In reality people don't just use one single model because it's simply not working like that
>>101056524>the emmapostingBeing reminded of some things I chose to forget
>>101056590>In reality people don't just use one single model because it's simply not working like thatHe never said that
>>101056524I hate being reminded of how easy it was to prompt properly big tits in the SD1.x models
post cursed gens
>>101056567Not sure you're exactly correct in the details.Still people clearly have reasons to stick with the older models and/or look at alternatives to SD3 for license, censorship and training mistakes reasons.
>>101056680Ok, if you're more the a lot is "good"-kind of guy: Except for very specialized SDXL models that really just do one thing well and destroyed the other trainings, they're all good all-around models
op says webui supports sd3... but the issues on github says otherwise...should I just use comfy?
>>101056950sure, use comfyui, it's good
https://arxiv.org/abs/2406.11831>easily surpasses state-of-the-art open-source models as well as mainstream closed-source commercial models including Stable Diffusion 3, DALL-E 3, and Midjourney V6.Based chinks who consider SD3 as closed source
>>101056366mobius is *supposedly* good and I've seen people get some incredible gens out of it but I personally haven't figured out how to get it to perform reliably. its trained to be a very beefy and robust model though and it might fit your vibe>>101056524looking forward to everyone picking debo gens unwittingly :3>>101056567>I heard that...interesting way to launder in unsubstantiated garbage as 'sourced'>>101056684just says that you should always find ways to appreciate what you have, even if its far from perfect. (except sd3)
is there a better way to include LoRAs in ComfyUI other than awkwardly chaining together the individual LoRA nodes with model and CLIP? And does the order in which you chain them matter?
>>101057056Yeah there's a number of different ones. I like rgthree's power lora loader.
>>101056567I 100% believe this
>>101057056I think most people use lora stacker nodes. IMO they should be the default loaders but that's not how it is rn.
>>101053907I know this gen... :)
>>101057047>deboAdvice ignored
>>101057056The Lora Stacker nodes I use most is from Efficiency Nodes BTW.
>>101056634I miss it. she's cute. better than the old dalle threads which were an unbroken wall of daisy ridley's goblin face.
where do I download a workflow for sd3civitas banned sd3
SD3 is powerful so long as you don't ask it to make a woman or a gun or anything violent or people not standing upright or sitting in a chair or
>>101057157im retarded sorry im reading op>NSA0
I will never be able to get proper swords on this piece of shit
>>101057157comfyworkflows.com
>>101057047>nobody wants him>waste his life anywaysImagine getting ran out of your own tranny discord
>>101057172>nooooooooo you don't understand!>the censorship! >the lobotomy!>the pruned dataset! >the license!don't care
>>101057078>>101057094>>101057120>lora stackerthanks, I didn't know about that.
>>101057274>anime fanservice if it real
>>101057274that top is crafted from the finest unobtainium
>>101053907That's generated? Believable good to me if so.Have a nice thread.
>>101057172SD3 is bad at guns, don't get me wrong, but only in the sense that SD was always bad at gunsI'm cooking up some terrible gun designs as we speak
Which thread has higher gen quality? /sdg/ or /ldg/?
what were the optimal values for XL again?scale and neg_scale in particular>picrel
>>101057285Don't get me wrong, I do like SD3.
>>101057551a hyperdimensional van
whoever said SD3 can't do guns was so full of shit. maybe it can't get particular guns right but if you want to dream up some disgusting abomination SD3 is exactly what you wantthis gun uses the mag as a grip. seems plausible to me (call of duty player)
>>101057706I guess having the mag grip shaped and doubling as grip could kinda work
>n-no ATF these are not printed
only the brains trust in bongistan would consider women haram but gore and bullpups okay
>>101057551Very funny image, I raffed hard.
The look of satisfaction
which notable poster of all time do you miss the most?
>>101058045petra
>>101058080slightly better gens out of the ti
how does civitai make any money?
>>101058111they don't right now lol
>>101058111>how does civitai make any money?Can they sell your IP address to vulture copyright lawyers?
>>101058045Smeagol
>>101058045debo
>>101058045That one anon.
>>101058217me too, that anon was the best
>>101058217I miss them like you wouldn't believe
New youth outreach program by the Catholic Church.
>>101055322>realistic lora>ponyYou might have better luck with non-Pony finetunes. Pony removed a lot of the realism from SDXL.
>>101054792SD3's l and g clips are the same as the ones used for SDXL. This guy is doing pointless rituals.
>>101058350SD3 has different G.O.O.N. parameters though.
howdy
>its that time again
my asshole is leaking and my tummy hurts
>>101058499Just Trani things
>d***
>>101058398:|
>>101058464Nice, just little uncanny
>>101054096I would kill for that.
>>101058717thxisnt everything in ai a bit uncanny?>>101058464
>>101058717jesus man. That face is pure crazy. If see someone smiling at you with eyes that look like that, you should probably run as fast as you can
>>101058738indeed.>>101058754kek, I probably would
Vanilla XL is not good at this.
>>101058813All I have is a girl and a robot while you have an entire village gathering their food for the winter.Nice!
>>101058775If i give you a dataset of pictures would you make an XL LoRA for me?
>>101054308>>101054267>>101054143Ppppromts?
>>101058842Ask another anon. I only worked on embedding for 1.5.
>>101058717>>101058775>>101058754but the crazy eyes are what draws me in ;_;>>101058738
>>101058813I think that's it. The greatest gen ever. I was there.
>>101058866Cool, is it true that embeddings are better for photorealism and real people?
>>101058953not him but I suspect embeddings would be better for the simple reason that SD has a ton of real human data it learned on already and just needs to know how to understand your instructions better; the gains from LoRA training would be outweighed by the loss of model capability.But it might depend on the person and how well-captured they are by the training data (and therefore how well they can be 'described' in conditionings to the model)
>>101058322I just wanted play with the level of prompt control pony has. i'm also super unfamiliar with sdxl models. the hidden gems are always the best
>>101058953Probably. Did a few from cosplayers.
Power Quokka
>>101058925
>>101058838She has no hips. That's a boy. You want to fuck a twink.
>>101058217I saw him a while ago and waved and he just ignored me
>>101059174a hole is a hole
>>101057430Yeah, that's one of mine, from last thread.Generating chibis makes it a bit easier to generate convincing looking stuff since many of the usual problem areas (like hands) are a lot easier.
>>101059174not him but I think she has ok hips, not exaggerated ideal but believable woman hips.
>>101059059I think if you wanted Pony control over realistic photos you would want to curate and booru-tag an enormous amount of real porn and train it the way PonyXL was trainedMany realism models are in fact half-assed attempts at doing something like this, minus the booru tagging.
>>101059224Yes. I have tried to become a homosexual but failed miserably. I found a pedophilic in me instead.
>>101059224Everyone knows what a "cloaca" is.
>>101057513>unskilled hentai artist bootlegsvs>timeless artist kino samplinggee idk anon
I can't get cascade nor XL to recognize halo armor, it used to be possible on some 1.5 fine-tunes and i also remember making 512x512 capybaras on halo armor before.
>>101059307Try harder please... It's cool to be homosexual these days anyway - just imagine a hairy man's ass mhhmm.Otherwise /sdg/ might not be for you!
>>101059368quokka > cloaca
I'm going to bed.Good night, everyone.
>>101059460gn
>>101059427Holy hot!
>>101059485I've seen a lot of women in my day and I'm at least 65% sure they don't bend that way
>>101059109
>>101059427Fantastic, I'm going to do something similar tomorrow. I'll see if CinematicRedmond will gen anything.
>>101054538What was your question?
>>101059460Gn anon
How many images do you need for a style lora? Like the lower end.
>>101059493Nice.
>>101054538Why did Mr Brady thought it was more than a hunch? How did he knew that the group must somehow form a family? when was...? When was the day they all became... THE BRADY BUNCH?!?
>>101059591>"What's the deal with AI art?"
>1 hour after eating the shrooms>>101059505
>>101059598now I miss seinfeld AI :(
>>101059460night
>>101059598I love how this model makes Jerry look so inbred
>>101059540Welcome back, Master! Genuinely nice landscape.
>>101059628thanks. one of the few strong points I found in mobius was landscape gens. alien creatures werent bad either. their model card says clip skip 3 but I kinda think skip2 is better....
>>101059612>(Cohesium:1.5)One of the most powerful trigger words, try it and post the gen.
>>101059650Interesting, why would they need clip skip 3... unless it's something nefarious...
I love Ran
>>101059670the most nefarious of all the clip layers
>>101059614Is that mars?
cubism>>101059534thanks
>>101059530Lower amount of images leads to higher chance of overfitting. It's better to have a lot of images. You need slight variations of each image, which you can do through img2img or manual editing if you can do art.
I'm helping!
First time in days a thread reaches image limit before reply limit, yay.
>>101059718I was actually wondering how to overfit so thank you for this advice. I sometimes prefer overfit loras you can weigh down if need be.
>>101059717>cubechecks out
baker pls
>>101059785300 seconds
>>101056634I wonder if Emma ever learned about it. She is a globohomo feminist so probably not since it never went mainstream? Hard to believe considering how many Emma pictures we generated back then.
BAKER
Next thread>>101059885>>101059885>>101059885
>>101059887>>101059887>>101059887
>>101059888>>101059894fug
>>101059919I was playing online vidya
>>101054691it's just some liver, it's food, not guro
nice
r.i.p.