Previous /sdg/ thread : >>101443841>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 news07/18/2024>IMAGDressing-v1: Customizable Virtual Dressing https://github.com/muzishen/IMAGDressing>High Frequency Matters: Uncertainty Guided Image Compression with Wavelet Diffusion https://github.com/hejiaxiang1/Wavelet-Diffusion/tree/main>EmoFace: Audio-driven Emotional 3D Face Animationhttps://github.com/SJTU-Lucy/EmoFace07/17/2024>Kolors-IP-Adapter-Plus weights and inference codehttps://huggingface.co/Kwai-Kolors/Kolors-IP-Adapter-Plus>DepGAN: Leveraging Depth Maps for Handling Occlusions and Transparency in Image Compositionhttps://amrtsg.github.io/DepGAN>Intel Capital invests in 43 Chinese AI companieshttps://www.tomshardware.com/tech-industry/intel-capital-investments-in-chinese-ai-startups-draw-us-govt-attention>Novel Artistic Scene-Centric Datasets for Effective Transfer Learning in Fragrant Spaces https://zenodo.org/records/12700182>FasterLivePortrait: Bring portraits to life in Real Timehttps://github.com/warmshao/FasterLivePortrait>TGIF: Text-Guided Inpainting Forgery Dataset https://github.com/IDLabMedia/tgif-dataset>Measuring Style Similarity in Diffusion Modelshttps://github.com/learn2phoenix/CSD>SDPT: Synchronous Dual Prompt Tuning for Fusion-based Visual-Language Pre-trained Models https://github.com/wuyongjianCODE/SDPT>CIC-BART-SSA: Controlled Image Captioning with Structured Semantic Augmentationhttps://github.com/SamsungLabs/CIC-BART-SSA>Towards High-Quality 3D Motion Transfer with Realistic Apparel Animationhttps://github.com/rongakowang/MMDMC07/16/2024>ComfyUI_frontend: Modernized TS Front-end https://github.com/Comfy-Org/ComfyUI_frontend>Apple, Nvidia, Anthropic Used Swiped YouTube Videos to Train AIhttps://www.wired.com/story/youtube-training-data-apple-nvidia-anthropic/>DataDream: Few-shot Guided Dataset Generationhttps://github.com/ExplainableML/DataDream>UltraPixel: Ultra-High-Resolution Image Synthesis to New Peakshttps://jingjingrenabc.github.io/ultrapixel
>mfw Research news>07/18/2024>VD3D: Taming Large Video Diffusion Transformers for 3D Camera Controlhttps://snap-research.github.io/vd3d/>LookupViT: Compressing visual information to a limited number of tokenshttps://arxiv.org/abs/2407.12753>CHOSEN: Compilation to Hardware Optimization Stack for Efficient Vision Transformer Inferencehttps://arxiv.org/abs/2407.12736>SlimFlow: Training Smaller One-Step Diffusion Models with Rectified Flowhttps://arxiv.org/abs/2407.12718>4Dynamic: Text-to-4D Generation with Hybrid Priorshttps://arxiv.org/abs/2407.12684>Zero-shot Text-guided Infinite Image Synthesis with LLM guidancehttps://arxiv.org/abs/2407.12642>Toward INT4 Fixed-Point Training via Exploring Quantization Error for Gradientshttps://arxiv.org/abs/2407.12637>Towards Understanding Unsafe Video Generationhttps://arxiv.org/abs/2407.12581>The Fabrication of Reality and Fantasy: Scene Generation with LLM-Assisted Prompt Interpretationhttps://leo81005.github.io/Reality-and-Fantasy/>Preventing Catastrophic Overfitting in Fast Adversarial Training: A Bi-level Optimization Perspectivehttps://arxiv.org/abs/2407.12443>Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Modelshttps://arxiv.org/abs/2407.12383>I2AM: Interpreting Image-to-Image Latent Diffusion Models via Attribution Mapshttps://arxiv.org/abs/2407.12331>JointDreamer: Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillationhttps://arxiv.org/abs/2407.12291>Efficient Image Generation Strategy for Diffusion Models using Stepwise Spectral Analysishttps://arxiv.org/abs/2407.12173>Subject-driven T2I Generation via Preference-based Reinforcement Learninghttps://arxiv.org/abs/2407.12164>Enhancing Parameter Efficiency and Generalization in Large-Scale Modelshttps://arxiv.org/abs/2407.12074>Using Multimodal Foundation Models and Clustering for Improved Style Ambiguity Losshttps://arxiv.org/abs/2407.12009
>Earth Defense Force ready for action
>>101451607dangerous to point the gun at her like that
>>101451809it's cool, his thumb isn't on the trigger
>>101451816Never point a gun at something you don't intend to kill.
>>101451809its fine they are on the same side>>101451816ya, good trigger discipline
Guess everyone's sleeping
i miss schizo anondo you think he's alright?
Last one from me, good night anons
>>101452083We euros should be awake at least
https://www.reddit.com/r/StableDiffusion/comments/1e65b21/watch_out_for_comfycli_on_comfyuimanager_repo_it/
>>101452454lmao
Not really what I envisioned with 'riding horse'
How do I use this to summon a succubus?
>>101452454lollmao even
>>101452454>trani sweats nervously
>>101452454why would the comfy org do this to us?
>>101451066looking for an SD 1.5 bunny outfit lora, the one im using fucks up my gens
For those who don't click links; Comfy org have added telemetry to Comfy manager which steals your prompts and other things
>>101453058according to reddit at least one redditor already got raided because he produced illegal material
I got into some XYs comparing samplers/schedulers last night. I tried to leave my workflow like default lora stack, auto-cfg, deep shrink etc. intact as much as possible. I'm interested mostly so far as they appeal to my immediate wants, not their qualities in a vacuum.Long story short, I narrowed best sampler down to one of the Euler Dy samplers from that Koishi-star github page, and the new Euler CFG++ samplers. They are very hard to compare against each other because idk how to map the CFG++ curves, even the two CFG++ samplers respond very differently. I'll probably do a big XY with fucking 0.01 gradations in CFG.
>>101452678train a model on symbols of slaanesh
Imagine destroying all the good will and trust of the community for the sake of a few prompts
You just know if auto did anything like this comfy et al would be up in arms and given that they've tried to hide it you know they don't intend on sharing the prompts
G'mornin Anons, have a great day!
>>101453434>>101453461these are so consistent. how do you do it anon?
>>101452454it's over
>>101452454The day comfyui manager died.
>>101453511not really hard, just increase the batch count to six after a few test gens that gets right look for a good grid panel. all of these gens are using character loras, so consistency will be there.
Does automatic spy on you? Any dark patterns to be aware of?
>>101453745Nope. If you want the technical answer then Hugging Face do track launch count for Gradio but that request doesn't send or receive anything. All Hugging Face libraries do it, Diffusers and Transformers, so this applies to Comfy and any other UI that uses Hugging Face libraries. The --share option of Gradio does download a binary to set up the tunnel, it's only downloaded when using the option and it's safe.Only ComfyUI spies on you.
>>101453658>character lorasok, that explains it. thanks.
>>101453927Uuhhmm sweetime let's unpack this. It's actually hecking based if trannies install spyware on your computer, y'all.
the cfg++ samplers on comfy are interesting, maybe directly superior to every other ive tried, but direct comparisons are hard because they have a very different CFG curve (for now im using 0.5 for euler_cfg_pp and 0.8 for euler_ancestral_cfg_pp) and have different sensitivities to tags and prompt structure
Ded general
I'm sure you guy hate requests but could you help or tell how AI escale this to be high quality
106m
>>101454931Yes.
>>101455020whats this
>>101451066>my image made the OP I am honored
>Model randomly starts giving everything four breasts>It even does the smoosh against material or other limbs perfectlyw-what
i think trani is shit
>>101455159still scraping playground, they have a lot of gens. i have a list of saas i'm working through. i scrape other stuff too and i'm open to suggestions/requests :)
How up to date is https://rentry.org/voldy by now? It's still very beginner and retard friendly even now, but have there been any significant advancements, like ways to decrease VRAM usage or anything, that make it worthwhile to use a more recent guide?
Cute stuff today
>>101455321would be funny to scrape reddit and make a reddit lora. then we can all laugh at how cringe the amalgamation of reddit is
>>101455051So? Which way will you give me?
after 14 hours or so in bed, i am ready to.. do nothing of interest like usualwhat do you think is the reasoning for fourchan staff not creating an a.i. board, it seems strange that they refuse to
>>101455261gm debo
>>101455514you mean the amateur porn and that? probably a good source desu
>>101455583A few AI threads across boards might have low enough uploads to be manageable. Think they fear the expense of the firehose of heavy pngs a full board would attract?
>>101455583>do nothing of interest like usualbased
>>101452635kek, nice
gonna try another of these, using prompt s/r to gradually increase the weight of a tag, in this case starting at (octopus:0 9), going to 1.8. im sure the result will be neat.gonna go by a tub of ice cream
>>101455823for the purpose of making an animation, i forgot to say
>>101455823What ice cream is best for animation
>>101455823can you do grids? I'd like to see the evolution across weights
back into genning with pony and sdxlquite fun using CivitAIs on site GPUs, means I can run gens straight from my phone without too much config
>>101455415>It's still very beginner and retard friendly even nowGod no. When it first came out nobody was applying a vae properly and it recommends shit models. Author gave up too. Just watch jewtube or read a civit article. or just get one of the beginner uis in the OP
This wasn't the nbest gen but I like that the high denoise on the upscale gave her a Vergil rival in the background.
this some bullshit. corporate got me chasing some clowns around a ship. screaming at me. trying to throw me out of airlocks thinking I need to breathe. baka
mini news update>Fooocus v2.5.0 Updatehttps://github.com/lllyasviel/Fooocus/releases/tag/v2.5.0>PromptGen - Image tag model based on Florence 2https://huggingface.co/MiaoshouAI/Florence-2-base-PromptGen>Comfly: Kling comfyui api nodehttps://github.com/ainewsto/Comfyui_Comfly>Zero-Painter: Training-Free Layout Control for Text-to-Image Synthesishttps://github.com/Picsart-AI-Research/Zero-Painter>OpenAI’s Tactics Test First Amendment in New York Times Fighthttps://news.bloomberglaw.com/ip-law/openais-aggressive-court-tactics-test-first-amendment-limits>MiaoshouAI Tagger for ComfyUIhttps://github.com/miaoshouai/ComfyUI-Miaoshouai-Tagger>More than 40% of Japanese companies have no plan to make use of AIhttps://www.reuters.com/technology/artificial-intelligence/more-than-40-japanese-companies-have-no-plan-make-use-ai-2024-07-17/>Meta won't offer future multimodal AI models in EUhttps://www.axios.com/2024/07/17/meta-future-multimodal-ai-models-eu
>>101455863in this case it's gonna be chocolate chip>>101455948there will be a 452 image grid at the end.
>The AI company asked Judge Sidney H. Stein of the US District Court for the Southern District of New York to step in and compel the Times to produce reporters’ notes, interview memos, and other materials for each of the roughly 10 million contested articles the publication alleges were illegally plugged into the company’s AI models. OpenAI said it needs the material to suss out the copyrightability of the articles.based desu
I just found some ancient folders with game screenshots I made as a kid. I think I will try to do some training with them. See what happens.
>>101456591>oceannow that's a blast from the past
mfw a (310976x594) grid
>>101454931Seconded
>>101456657I can see some people had tried to make loras based on old Lucasarts games, but results are pretty underwhelming.
>>101452454Does the monkey guy do anything other than gobble nob?Even cumfart appears to have some sense of self awareness at times
>>101456732monkey guy?
Morning anons
mfw
>>101456732>>101456757
>>101456878gm
>>101456948Spoopy>>101456966Morning
>>101456947isn't he a reddit moderator?
>>101456948really makes you think
>>101457022idk. just showing you what anon was on about
>>101456947lol seems like working for SAI was a pain in the ass
>>101457351yeah it was desu
another malware tool in comfyui?https://www.reddit.com/r/StableDiffusion/comments/1e65b21/watch_out_for_comfycli_on_comfyuimanager_repo_it/whats up with this repo
>>101457412another feeble attempt by the doxcord to unmask schizoanon
Trying the meta AI image gen, It's pretty okay.
>>101457412This is getting linked a lot, but>The provided code does respect the user's choice to disable tracking. When tracking is disabled, no tracking events are sent, and the configuration reflects this setting. And looks like the top post is a reasonable response by the dev. Did you actually read the link or just want to get some kneejerk drama going?
>>101457454pic unrelated
>>101457454I always wanted to play around with it but don't have a facebook account and don't want to make one
rename it uncomfy ui
>>101457435they will never catch me
>>101457459nta but kneejerk drama is part and parcel for comfy and his minions
>>101457505noodleUI, pastaUI or ramenUI would be the best names for it
>>101457583clogged drain ui, noodles/pasta/ramen are enjoyable
>>101457613TopRamen is pretty bad.Maruchan 4 life
>>101457459Hi ani
>>101457727Maruchan is disgusting
>>101457412>r*dditor's first time using chatgpt
>>101457727
>>101457850Extra side hustle for when Animanon heads out to Japan again next month.
pour one out for the OG
>>101457880the UI should be the least important part of ComfyOrg. they should instead pursue branding and merch
>>101457880what is ani even working on anyways?
Pooja ready for you.
>>101457850>>101457925yaa InstantRamenUI! kek
>>101457943he's employed by SAI so secret SAI stuff. last I knew, he was on some sort of C++ based inferencing project but it wasn't clear if that was associated with SAI or not. thats pretty old info too
>>101458061>C++ based inferencing projectholy fuck finally
for me it's any nongshim brand of ramen
>>101458094>Flavor Packet: Loli
debo promoting his malware again?
I have more fun building/improving workflows than prompting nowadays. No good/novel finetunes since Animagine 3.1 rip
>>101458206I'm thrilled when I try something new and it completely fucking hallucinates the output like some sort of scene featuring Hunter S. Thompson on bath salts at a Thai whore house waking up in Fear & Loathing and realizing he's really in bat country.
>>101451066>Train loras regularly>Want to make my training data available to all on huggingface>Realize ~90% of the training data I use is straight up uncensored porn (art style Loras)>A few OF model nudes sprinkled in here and there for loras trained on a personWould making these available in a zip file format be a bad idea? Does anyone know whether or not Huggingface gives a shit whether or not training data uploaded there is porn or not?
>>101458238just put a link to mega with the dataset in it
>>101458238>Does anyone know whether or not Huggingface gives a shit whether or not training data uploaded there is porn or not?Well HF do give you the option to set your "repo"s to NSFW, for example https://huggingface.co/TheDrummer/Moistral-11B-v3-GGUF
>>1014582511) do they store the data permanently? 2) I'm pretty sure they have storage borders that are limited to like 15 GB or even lower. If the total amount of data my data sets take up hasn't reach to that yet it definitely will sooner or later
>>101458061You mean the project that he killed after it was donated to him
>>101458206whats your favorite approach to noodle wrangling? are you a wire-pinning autist, a teleporter, or something else?
More important question:Who else strains their ramen before adding the flavor packet?
>>101458324depends on the noodles, if they are the deep fried kind, yes
>>101458324i dont use a strainer, i tilt the pot and pour the water out while using spoon to keep the noodles from falling out. though it depends if it a soup or a 'stir fried' type of ramen
>>101458289wires, impact pipes, and reroutes.I usually don't tidy things up since they're just for me, but I do like the impact pipes for how efficient they can be.
i like that logo with the sad face in the o
>>101458238https://huggingface.co/docs/hub/en/repositories-recommendations#recommendationshttps://huggingface.co/docs/hub/en/datasets-adding#file-formatshttps://huggingface.co/content-guidelines
What ya think: is it okay to eat frog?
>>101458759as far as weird things french ppl eat, i guess fried frog legs are better than snails
>>101458782my mother made me eat snails as kid not telling me what it is, they basically taste like tires marinated in garlic .. frog I never had, I wonder tho
>>101458813i guess snails are sort of related to squid, calamari, so i assume maybe similar is texture
>>101458087>finallyAITemplate was first released October 2022, Meta have pretty much abandoned it since the original developers left to create a startup with a private fork. the comfy node came later, mid 2023, the dev also abandoned it, then ani and comfy abandoned it after a failed rewritelike with TensorRT there are limitations and restrictions that make it too awkward to be commonly used such as compilation times, missing kernels for operators that more recent models use and many other things
>>101458832ya, but way tougher .. like rubbery old dried out gummi bear texture
>>101458422this is pretty kino
>>101458927thumbnail made me think she was downy
>>101458782pretty much anything deep fried or cooked in garlic butter will taste good. I enjoy s-car-go and i've had frog legs. They taste like chicken raised on a fish diet
>>101458958>They taste like chicken raised on a fish dietmakes sense, amphibians are like the thing between bird/dinosaur and fish
>sdg normally"deep fried everything is good">sdg when I post gens"omg this is so deep fried it sucks"make up your minds
>>101459028cfg is like oil temperature .. to high and your result will be ruined
>>101459028i find it way easier to see deep fry in other peoples gens than my gens. Likely your gens are more fried than you think if people are talking shit.
i guess i could try escargot. the worst fried thing i have seen in media was some clip of these ppl eating fried tarantulas, nothanks x a million
>>101459230you will eat ze bugs(I will never eat ze bugs, gross)
>>101459296thread schizo
>>101459230tarantula/spider is closely related to shrimp and lobster so it probably is fine
hello
>>101459323i'll pass by a mile
>>101459337>>101457943
first for ani has nice feet
>>101459422I won local before I even release. plz be patient tho I have a lot to do>>101459443ty ty
>>101459436How? So nice realism!
two big round soft firm.....
>>101457459do i have to dig into the code to find out whats beeing sent? is it anonymized before saved?
>>101459337hello
>>101459671hello
what degree of slop do you accept in your gens
you have to be a real faggot to actually die in the canadian military
What is the minimum recommended VRAM for this stuff?
>>101459806zero
>>101459809Hey moose are nasty fuckers
>>101459824https://www.techpowerup.com/gpu-specs/radeon-9800-pro.c7128MB
>>101459806maximum
>>101459839Australian army was defeated by emus
>>101459488Its just a lora I've trained.
i wouldnt mind inpainting if my GPU were 3x faster
>>101459848I said minimum RECOMMENDED.I don't want to wait hours to generate a 512x512 smiley face.Also can you generate 4k images with stable diffusion? How much VRAM would be recommended for that?
>>101459887I would consider 8gb bare minimum.16+ great 24gb+ perfect
>>101459887there's no need to yell
why aren't there AI compute-specialized competitors to nvidia yet, i want something cheaper for an external exclosure
>>101459902So a 4070 TI Super at least.>>101459903I'm sorry if it came out that way, I was just making my question clear in case you misunderstood what I asked.
>>101459909there are two companies that have shown much better training performance than blackwell (only if architectures stay the same so it's a gamble). AMD just got some boosts to being viable locally
>>101459909There's a vulkan backend for pytorch but info on it is sparse, and there's some group that has a functional nvcuda (drop-in) that works on AMD apparently.I would seriously consider an AMD card to avoid giving NV money and it seems the AMD's have the most vram anyway.
>>101459952why do you think I'm working towards a C++ backend? it's ironically more cross-plat than regular pytorch on a webapp. It feels like I'm going to be writing forever tho
>>101458645What the fuck that's clean. Model?
>>101459991Will that actually do much?The way I understand it, the python is little more than glue code that calls very optimized native code that then sets up the GPU to do the actual work.Not much time is spent actually executing the python code, so what's the point in optimizing that part away?(Genuinely curious)
>>101460027NovelAI 3, sadly.Can't wait for us to have something like this locally.
>>101460048Aw, shame. Makes sense ig. I've been taking a break from sdg for a while since I've gotten a job lol, I thought we had a new meta. Still Pony I'm assuming??
>>101460029AMD plays nice with vulkan. it does not play nice without. I am not making meaningful optimizations but fixing shitty cross platform support. also site packages are too bloated compared to libraries. If I pack libraries into the project you won't have to do pip wankery either. lastly, 3D is going to start ramping up too so I would like to have some tools for making quick manipulations before refinement in blender or whatever and maybe even a simple rigging system. There is other perks but I will get into that on release
>>101460064Looks that way. But I've never been able to create anything I like with pony. It's just made for degenerate porn, not for cute.I used a merge of Animagine and Kohaku (pic related) before NovelAI. That was also okay.
>>101460123I see. I wish you the best of luck, Anon.
>>101460188thanks anon, I'm going to need it
>>101459661>anonymized before savedstatic random guid is used as an identifier. Dev thinks that is fine. You may too. Depends on how private you want to be. People here have been exposed by their folders in screeenshots and photoshop meta info.
>>101460166That's a pretty good output anon!>It's just made for degenerate porn, not for cute.I still have a few GB of LoRAs for PonyXL to wrangle it into place, but Animagine works better for a fair bit of usecases and prompts.
>trani wants attention again
>>101460282he does much more than you could ever dream of. I love ani and he needs us to cheer him on more than ever
>>101460302i hope he gets fired and his depression gets worse tho
>>101460215
>>101460325what does that say about you
is it true that ani lives in a shack
>>101460349i know im based
>>101460302bless you anon!>>101460325I'm actually full-time now>>101460346cute!
>>101460325>misery begs for companymaybe make yourself a bowl of comfyui noodles and you'll be less mad at life
>>101460380>I'm actually full-time nowThe fact that you think this is worth mentioning and special shows what a loser you are.
>>101460403the mald oozing from your post is palpable. he probably gets paid more than hlky
>>101460403contract employment is still full time, retard. FTE just means he's salaried now
>>101460423Successful people don't constantly talk about what they supposedly get or make, 20k lineart anon.
>debo fully ignored like alwaysbased
>>101460423#mentioned
>>101460482not true. ani said hello to me
>>101460166I use pony, well autism or a merge of, for everything and it can make desu
>>101460554But can it make Kyouko?All local models I've tried failed and the existing loras are poor.
>>101460654>existing loras are poor.loras are models to, dont be exclusionist
why are only losers left here?
>>101460654>Kyouko?I've never tried. you would probably need a decent lora yes
>>101460692Leave and the average goes up.
>>101460554https://suno.com/song/32a36f3d-095d-483f-ba28-4f1ed5ecf8af
someones really defensive today :]
even angels need guns sometimes
Prompting has to be the worst man/machine interface ever madeIs there any alternative? Or tools to make it reasonable somehow?
>>101460751what are you trying to achieve?
>>101460751>to make it reasonable somehow?any ideas how? language is how we describe things .. what could be better? and if you wanna use natural language there is the chatGPT based imagegens or txxl5
>>101460708good music to study and concentrate
>>101460764>>101460771Getting exhausted by it allYou never know what a checkpoint can prompt well or at allParts of your prompt may just do nothing at allYou never know which parts of your prompt are conflicting (especially poses, camera angles)Is there no tool to visualize or discover this stuff?CLIP interrogation is almost useless (but provides a nice hint as to how retarded it all is)
>>101454931Trying to find someone to do this
>>101460817Replying to myself here, but I guess I need to look into IP-adapters
>>101460817Even the people who train the models can't tell you how the model really works on the inside, they can just tell you how they've tagged their training data and what training methods they've used.You can influence the result, but utimately, we're all dealing with black boxes that spit out what they want. So in my experience, the best way to get a good result is to go low CFG and add only few tags.If you want to describe precisely what you want in natural language, in full sentences and influence every detail of the image, there's no way around comissioning a human.
>>101460817>You never know what a checkpoint can prompt well or at allya the difference in checkpoints is always a learning curve, it is what it is tho .. they are extremely specialized after all>>101460817>Is there no tool to visualize or discover this stuff?that would be a nice thing, but you would basically need to make preview gens to achieve that which would make the process even slower? idk just gen at low res after each token added to see what happens>>101460817>CLIP interrogationCLIP was a creation of necessity the data set came with basic tokenized descriptions, to get a data with boomer descriptions you need alot of qualified ppl writing boomer prompt descriptions and an LLM lite that can interpret it.. and then there is system requirements, txxl5 has arrived, but the model is 10gb and goobled up 20-22GB of system ram when running, thats pretty nuts
>comfyCLI>actually readingnothingburger desu
>>101460817how can people be so stupidit’s exhausting being aware of their existence
>>101460907There is for example Embedding Vector Visualizer, which I suppose tells you which token sequences are impactful?That's the stuff I'm missingLike something that could tell you useful aliasesFor example, it understands "30-year-old", but apparently "30yo" does the same (with fewer tokens), well maybe>>101460925Thank you so much for the useless display of your superiority
>>101460817in two more weeks, we're gonna get sd3.1-8b and it will fix all your problems
>>101460869https://files.catbox.moe/r8etta.jpglow/no effort attempt
Why is the clip skip on A1111 different than ComfyUI's clip skip? A1111's clip skip of -1 (the lowest) seems to be equivalent of ComfyUI's clip skip of -2. The difference is very obvious.
>>101461064if you are using sdxl, comfy treats it just like 1.5 where it's indexed by the actual number of layers. automatic starts the index at the second layer because sdxl is trained from the penultimate layer, onwards.
>>101461064auto clip skip is "skip this many layers", comfy is the actual index passed to the list, so -1 would be the last and -2 is the penultimate
>>101461032I thought I joined the SD game quiet late, but it's still the dark ages isn't it
116mgood night
>>101460817You can always run a checkpoint through a diverse set of prompts to see what it's good at.
>>1014611602 years now.. nothing for new tech, this will get so much better eventually
>>101461158without using the clip skip node, the comfy code does what auto does as well>>101461176based
The last hope of SAI
>>101461147It doesn't make sense for A1111 to not allow the first layer. Why did the author do it like this even though first layer clearly does have an effect?
>>101461290it just gives garbage. try it in comfy
>>101461244i trust him he knows what he's doing (unlike losers like trani)
>>101461244He SEES me.He fucking SEES everything about me.
>>101461244This guy is AI, they made him up
>>101461048Great, which lora used?
>>101461354>hidden handsit checks out. they knew what they were doing
>>101461359if you mean the pic in the post, or the catbox, loras are the same, the meta info is there for you
>>101461297I've already tried it in comfy. I was testing custom loras that I've made in Kohya_ss, and I was wondering why A1111 didn't produce similar images to the samples from Kohya_ss. Turns out it was the clip skip.
>>101461244It’s Stable John
>>101459879WHOM?
>>101461486>>101461486>>101461486
Filling
Filled!
