Previous /sdg/ thread : >>101827381>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>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>Black Forest Labs: Fluxhttps://huggingface.co/black-forest-labs/FLUX.1-schnellhttps://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellhttps://comfyanonymous.github.io/ComfyUI_examples/flux>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
First for Debo is the thread schizo and trani a confirmed pedo
>mfw Resource news08/11/2024>SimpleTuner v0.9.8.1 - much-improved flux traininghttps://github.com/bghira/SimpleTuner/releases/tag/v0.9.8.1>Forge - BitsandBytes Guidelines and Flux [NB4]https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/981>ComfyUI_bitsandbytes_NF4https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF408/10/2024>Forge 2.0.0 https://github.com/lllyasviel/stable-diffusion-webui-forge/commit/cfa5242>X-Labs Flux LoRA Collectionhttps://huggingface.co/XLabs-AI/flux-lora-collectionhttps://huggingface.co/Kijai/flux-loras-comfyui/tree/main/xlabs>SwarmUI 0.9.2 Beta: Flux, SAM2, UI Improvementshttps://github.com/mcmonkeyprojects/SwarmUI/releases/tag/0.9.2-Beta>ComfyUI v0.0.5 Flux improvements and Hunyuan DiT controlnetshttps://github.com/comfyanonymous/ComfyUI/releases/tag/v0.0.508/09/2024>ComfyUI-Veevee: Video2Video Framework for ComfyUIhttps://github.com/logtd/ComfyUI-Veevee>BIGstockimage2M: ~2M stock images and captions @ ~0.4MPhttps://huggingface.co/datasets/bigdata-pw/BIGstockimage2M>Pitchfork Dataset: ~14 thousand album reviews and scoreshttps://huggingface.co/datasets/bigdata-pw/Pitchfork>ComfyUI-LLaVA-OneVisionhttps://github.com/kijai/ComfyUI-LLaVA-OneVision>AMD asks developer to take down ZLUDA, dev vows to rebuildhttps://www.tomshardware.com/pc-components/gpus/amd-asks-developer-to-take-down-open-source-zluda-dev-vows-to-rebuild-his-project>Freeway Animation HunYuan Demo: First large-scale Hunyuan-DiT fine-tuninghttps://huggingface.co/Laxhar/Freeway_Animation_HunYuan_Demo>DreamBooth training example for FLUX.1 [dev]https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/README_flux.md>ComfyUI on an On-Premise Kubernetes Clusterhttps://github.com/Curt-Park/comfyui-onprem-k8s>pyBregMan: Python package for Bregman Manifoldshttps://franknielsen.github.io/pyBregMan/index.html08/08/2024>ml_mdm - Matryoshka Diffusion Modelshttps://github.com/apple/ml-mdm
>mfw Research news08/11/2024>JambaTalk: Speech-Driven 3D Talking Head Generation Based on Hybrid Transformer-Mamba Language Modelhttps://arxiv.org/abs/2408.01627>SAT3D: Image-driven Semantic Attribute Transfer in 3Dhttps://arxiv.org/abs/2408.01664>GLDiTalker: Speech-Driven 3D Facial Animation with Graph Latent Diffusion Transformerhttps://arxiv.org/abs/2408.01826>Latent-INR: A Flexible Framework for Implicit Representations of Videos with Discriminative Semanticshttps://arxiv.org/abs/2408.02672>Working with Color: How Color Quantization Can Aid Researchers of Problematic Informationhttps://arxiv.org/abs/2408.03424>Decoding Visual Sentiment of Political Imageryhttps://arxiv.org/abs/2408.04103>Rethinking Visual Content Refinement in Low-Shot CLIP Adaptationhttps://arxiv.org/abs/2407.1411708/10/2024>iControl3D: An Interactive System for Controllable 3D Scene Generationhttps://arxiv.org/abs/2408.01678>A General Ambiguity Model for Binary Edge Images with Edge Tracing and its Implementationhttps://arxiv.org/abs/2408.01712>TS-SAM: Fine-Tuning Segment-Anything Model for Downstream Taskshttps://arxiv.org/abs/2408.01835>DeMansia: Mamba Never Forgets Any Tokenshttps://arxiv.org/abs/2408.01986>On Using Quasirandom Sequences in Machine Learning for Model Weight Initializationhttps://arxiv.org/abs/2408.02654>Decorrelating Structure via Adapters Makes Ensemble Learning Practical for Semi-supervised Learninghttps://arxiv.org/abs/2408.04150>LogogramNLP: Comparing Visual and Textual Representations of Ancient Logographic Writing Systems for NLPhttps://arxiv.org/abs/2408.0462808/09/2024>Puppet-Master: Scaling Interactive Video Generation as a Motion Prior for Part-Level Dynamicshttps://vgg-puppetmaster.github.io/>Enhanced Prototypical Part Network (EPPNet) For Explainable Image Classification Via Prototypeshttps://arxiv.org/abs/2408.04606>Img-Diff: Contrastive Data Synthesis for Multimodal LLMshttps://arxiv.org/abs/2408.04594
Reminder: >>101833846>https://imgsli.com/Mjg2NTE5>nf4 is closer to my prompt, which should be a live action aka real person
PSA: NF4 is the better model format for gens. It has noticeable improvements beyond FP8 and even FP16 generations. You should be using NF4.
>>101834483Good night/morning, anons! I'll see you later on before I leave :]Hopefully the angry imglets aren't around but hey, they have nothing better to do LOLIt was fun while it lasted!
>>101834533This was a test on a single seed, Anon. I haven't gotten many fp16 examples because of the slowness of gen. This doesn't really prove anything except that on seed 0 for that particular prompt nf4 was closer to it,Don't ever use my examples again unless you know what you're taking about.
>>101834536Based illyasviel saving us from comfy homosexuality
>>101834546Take it easy dude!
>>101834546hellogoodbye
>>101834546Pedo apologist.
>>101834548I know that every one of your examples has shown nf4 to be the clear winner, faggot. Don't blame me if you need to gen 50 more seeds before coming to the clear conclusion.
>>101834553retard.im saying if you cared so much about the purity of a gen in it's original intended format, then you wouldn't be adding LoRAs to them like everyone here did with SDXL and Pony. NF4 is simply another tool to enhance your gens, much like one of the many SDXL model derivatives or LoRAs. only reason you have opposition to NF4's clear superiority, is that it shows that Ilya was the superior gay programmer of the couple.
>>101834591every one of my 2 examples?
>>101834668You've either posted 4 nf4 examples or another anon is genning the exact same subject matter and comparisons concurrently. The point I'm making is that nf4 is doing great in spite of trani's meltdown, stop being an autistic retard Jesus Christ.
You were a lot tougher half a year ago thread schizo.Now go ahead and cry for help at jannies.
>>101834717I posted some from forge earlier but it's not a good comparison because forge and comfy produce different gens. Cats, functionally identical https://imgsli.com/Mjg2NTI3
>>101834722shut up shut up shut up!!
>>101834742>that pictureshe's literally me!
>>101834741Wow, it's almost like it's all exactly what I'm saying: nf4 is great and the schizo pissing and sharting himself over it being le bad is wrong. Now fuck off with your "don't quote my examples" bullshit.
>>101834660adding loras? i dont like loras i only gen without them because i dont even like them as a concept. i dont know who you think i am but i havent been around here since automatic came out. I'm opposed because i use llms and it's well known that you degrade quality by doing this. ai img users dont know this and are vulnerable to lies about it. if people actually believe they're equivalent that hurts ai img community as a whole.
>>101834784You are literally the schizo, Anon.
>>101834792>it was fight club all along
>avatarfags are a great addition to /sdg/You could have avoided all of this anon
I haven't been here in a long time. Is there a place where I can download flux pro?
>>101834790i don't care who you are, and nobody else does here. what you say has no value, and nobody here wants to be listening to you.
Did anything happen? Why are Julien and Debo in melty-mode again?
>>101834815flux pro is the saas model. its not open
>>101834418>miku will never bounce her hips back and forth while smiling sweetly as she listens to how NF4 gens are objectively superiorwhy live
>>101834759When's the last time he posted?
>>101834861need to lick her tummy
>>101834815can grab the dev version here for examplehttps://huggingface.co/ChuckMcSneed/FLUX.1-dev/discussions/1
>>101834854 oh that explains it>>101834908I'll check it outthanks ni/g/gas
Ouch! Look what you did...
>>101834852illya revived forge to make nf4 flux with superior gens, and comfy seethed, claimed nf4 was shit, then swiftly copy + pasted illya's code (literally). so they're on damage control duty as comfy's cuckslaves
>flux1-dev-bnb-nf4.safetensors>11.5 GB
>>101835259Does this mean I can run flux locally without downloading 30gb worth of shit or does the clip etc have retarded file sizes?
>>101835296No it's using an fp8 clip I believe, it all runs in vram for me.
>>101835453Promising I wonder if further optimization will be possible, still a lot better than where it started
never found dinosaurs very interesting
Why is comfy always so petty and how have more people not noticed? I'm surprised plebbit doesn't hate him aren't they all about cancel culture when someone's a perpetual turd
"What is that THING doing on my back!?"
>>101836103I'm just a poor nigga from the streets generating ai pics for fun, I don't care about any of that
>>101836103He was deeply in love but got eventually rejected... This could be anyone.
>>101836103Actually reddit is laughing at cumfart and Julien whenever they show up
>>101836103he's the only one not using shit frameworks or code
>>101836355>just download these 120 custom nodes and execute them for my workflow, whats the worst that could happen? :3
>>101836392>what happens when I intentionally put the shotgun to my leghaha you sure showed him!
Did comfy really steal illyas codes?
>>101836520No
...
>>101836569nice
>>101836520Yes, shamelessly.
>>101836577No, Pony takes way too much wrangling and even when it's fine the background is most likely fucked up which needs to be manually corrected and then upscaling it takes even longer.Pony is great for making jokes.
>>101836614I didn't know, I apologize
>>101836627You should be sorry!
>>101836520Yeah, what a faggot after all the github drama with him and julien lmao
>>101836679who is Julien?
holy fuck, fp8 was 15s/it and nf4 is 2s/it
>>101836758frfr?
>>101836772no screenshot so it didn't happen
why is python folder so massive
>>101836826python is shitty bloat but you are stuck with it because everyone in the space is niggerlicious (happy terry day btw)
>>101836772massive difference for 12gb vram. I'm going to run nf4 model + fp16 t5 next and see what speeds look like
>>101836848Ty Anon, lost a good one.
>>101836848It's been 6 years already
>>101836848;_; i miss him
>>101836825>no screenshot so it didn't happenhere you go
>>101836900ad = ag, oops
best ai tool outside of sd for decensoring images?
ok, the fp16 t5 still doesn't fit>100s/it
Forgy or Comfe?
now nf4 is at 34s/it?? wtf? why would it change?
>>101836990If you want to be a bigot, ComfyUI is the right choice.
>ded schizo general
>>101837169bien
>>101837202It's a message to potential criminals...
>>101836990Gotta be honest. All the gens I tried in forgy were shittier than my comfy gens.It might be because I was using nf4 in forgy and fp16 in comfy, but I can't test the fp16 one in forgy so..
>>101837210What did she do?
>>101836990Forge did speed up generations from an avg 15s/it to 5s/it
>>101837228>she
>>101837228She downloaded a car.
>>101836990forge still uses shitty auto code. comfy is faster
>>101837272criminal scum!
>>101837280Comfy uses shitty forge code
howdynow on forgetook some time to get the dependencies right (and i think i just accidentally stumbled on the right combo) but it werks
>>101837386Use the same prompt but add >,ukiyoe, somewhere in between.
>>101837422faster or slower for you?
why are my speeds all over the place?
>>101837455on a 3090ti, i'd say it "feels" faster, at least non-hires upscale wiseit's still buggy tho, every other gen i have to copy/paste the text prompt back to itself or i get the "TypeError: 'NoneType' object is not iterable" error (i already removed xformers)otherwise, the lack of headache from fucking around comfyui is a major plus for me
>>101837481Memory bottleneck, I mean the system ram means a lot in this case too. 64 GB would be optimal.
>>101837481ran out of memory, 491 seconds.. dude. be glad your pc didn't combust
>>101837526>>101837537my ram is never maxed though
>>101837499ti bro, you able to get a decent upscale?
>>101837537My friend died this way and I'm not joking. He's passed away because his computer combusted.
It's not her fault she likes to have it all.
>>101837499this is 1280x1024 -> x1.5 hires upscale
I don't dare to post this here but I think this is hilarious. >https://files.catbox.moe/3lu09j.jpg
ok I guess grouping my model nodes impacts my gen speeds???? ungrouping the nodes is back down to 2s/itwhat the fuck comfy
>>101837574>>1018376641280x1024 -> 1.5x hires ups (1920x1536) -> 2x upscale -> 3840x3072
>>101837664yeah doesnt look too hot. might be the initial gen tho. do some a/b testing and see if the image gets darker after upscale, shouldnt happen. flux can do fantastic resampling with 0.2 ish denoise and around 12 steps>>101837714whee. big girl!
Afternoon anons, happy last day of the Olympics
>>101837429
>>101837751>darkeralso depends on upscaler model i guess
>>101837811Less pronounced than I thought. I often get visibly abstract paper-cut out images but I guess this depends of the model and other stuff in the prompt too.
>>101837864no no, this looks horribru. here made some slide with that pink nurse, 0.2 denoise (can always crank that up) https://imgsli.com/Mjg2NTU1
>>101837939it's more than likely my prompt
oldie, seems forever ago
>>101837978just lower guidance a bit if baked, 3.5 pushing it ever so often beyond the BAKE line
>>101838021https://imgsli.com/Mjg2NTU2before and after hires, using 4x_foolhardi_remacri for upscaler on hirescfg 3, steps 20, hires steps 15looks fine to me?
>>101834742smelly ugly femcel
>>101838088nice, looks very clean! try less steps for upscale, like 12, cfg 0.2, also maybe nmkd superscale (remacri can brighten image) or for a more agressive approach ultramix balanced.>>101837798international quokka, I like it
>>101838218>cfg 0.2,wot
>>101838293denoise, ahem. flux resample seems to work just fine with a pretty low denoise
Pony is much better without its horrific backgrounds.
do we know why the guidance goes up to 100 and kinda shits itself at 1.0?>picrel guidance 94.3
>>101838282prompt?
>even mint pussy moved to /ldg/the absolute state of affairs
>>101838386>Cute elf girl, sitting, desperate pose, full body view, wide angle, dynamic camera angle,>short messy blonde hair, blue eyes, wearing prison rags, bare feet, shackles, scratches, blushing, crying,>purple background, wide angle view, Basically listing everything what you think is worth listing, then wrangling the prompt based on the output.
you just know
>>101838418>>101258654
>>101838418>>101838503i check out /ldg/ for information, but i post gens on /sdg/i like the comfy speed on the threads here (no pun intended)
>>101838418you love avatarfags now?
>>101838585he's always been my biggest fan
80m
>>101838790I was thinking about downloading some of the stock photo stuff but I don't have the space right now.
Fuck Pony, it cannot even draw a straight line when it comes down to backgrounds.It's only popular because of the fetishists. Too bad the author thought so highly of himself and basically cancelled his next excreation.
>>101838808stock image set is only like 100gbnice thing about webdataset for training is that you can stream it, either directly or using the cache mode and a limit of how much to keepyou can also use any number of shards rather than all of itwith http its best to use cache in case of disconnectsbig training will typically have the webdataset on the cloud provider's object storage because the gpu nodes will have a fast, stable connection to it
>>101838974It's better to have it locally.I was thinking what is your access in terms of galleries and such? I know they don't post everything online.Would be cool to have a real art collection.i.e. 10 years ago I remember the Museum of Amsterdam posted selected artworks in high resolution based on their current exhibitions etc..
>>101836990Krita :) also Photoshop
All right I guess this sums it all.
>>101839025yeah for such a small set you might as well have it all locally, just wanted to explain the benefit of webdataset and like how training on huge sets like laion happens without needing to store everything on the nodesatm i've only done one museum/art gallery, Tate https://www.tate.org.uk/search?type=artwork https://huggingface.co/datasets/bigdata-pw/tate it's ~66k rather than 79k mentioned on tate.org.uk because some are unavailable due to copyrightin a lot of cases i can get more than what's on offer as standard, like finding the high resolution image when they only show a preview, still it has to be online somewheremore art is planned, maybe ill take a look tomorrowonly limits on what i can get is server capacity, time and coffee intake, my mongo server is a little overloaded atm, scrape nodes are a bit too
>>101839337I'm actually really interested because this is very cool. You are restoring and archiving some part of the internet for the future generations to see and so on.
This is work safe in my opinion.
>>101839375i appreciate your interest. i should upload to archive.org as well, usenet would be another option for long term retention, 10+ years from some providersit would also help in cases where i don't have the capacity to store it myself, like all the images from diffusion1bthe services will inevitably shut down then all that history and progression of ai generated images would be lost. i only have a (relatively) small part of the images downloaded, around 13tb, i had to stop the image download after a drive failed to avoid any potential further failures while resilvering and havent restarted it yet, i estimate the complete collection would be several hundred tb and atm my limit is another ~100tbcivit loras are another interesting part of ai history which is why i'm downloading them all
>>101839601Yeah because even place like cs.rin is not a piracy site but its moderators are actually autists and sort of librarians.At the moment I would love to help but my internet connection, while fiber, is really limited.It's just fascinating work.
>>101839601I presume that in ~5 years or so the amount of data explodes but it will be mostly trash.CivitAI as whole is not interesting at all or maybe as a snapshot in time.
>>101839598My boss disagrees, now I have a meeting with HR
>>101839601Thanks for doing your part in archiving
>>101839762Tell the HR you were drinking and instead of discriminating you for looking at comics they will give you a free vacation.
>>101839831Is this flux still?
>>101839831ya, filename shows what it is
>>101839627suggestions of sites are just as much help as anything else, like with art galleries/museums, if someone made a list of this site has that it would save me the time of doing it myself. or something unique like one suggestion was https://www.1001tracklists.com/ which curates radio playlists and so on, having curated lists of artists that go together will help with other projectsfor some projects like video game screenshots i'd like to set something up for people to contribute screenshots directly, like, put together a list of all the games for all the systems then people can pick a game, play it, take screenshots and submit them. there are of course sources for this like steam or taking frames from twitch videos, but the public contribution model works very well>>101839702with civit sometimes the creators have deleted the loras that people are using which is a shame>>101839772thanks for your support!
>>101839856Oh I didn't bother to hover over the filename. It's like way sharper vanilla xl plus extra features.One day when someone will train it further and stuff it'll be interesting.
>>101839868I mean it has the same certain features vanilla had in some cases.
>>101839862I don't by know most of the big museums are not that open on internet at least not for free.Sotheby's is one interesting place but I have no idea how they are regulating their listing (which change every few weeks).
>>101839911I don't know if *
Not as "16 bit" as I hoped but I'll take it. Can Flux make pixel art well?
Last one, I am sorry if.
>>101839868>>101839888i dont know about vanilla xl comparison, since it took a while for sdxl to get to something like this quality/prompt coherence and you still need a finetuned model and/or loras (rather than just the sdxl checkpoint)flux is more like sdxl 3 but not as gimped imo
>>101839984I was talking about the images - this applies to every AI image. Flux is sharper of course but it is not by any means 'smarter'. It might be smarter in terms of boomer prompting and associations.
"What a nice night."
>install Git, Python, and reForge>keep getting "Torch is unable to use GPU" error>GPU is NVIDIAI've already deleted the venv folder and re-ran, and even completely deleted and reinstalled the entire thing, but no luck. Am I missing something?
>>101840096Have you put Python and CUDA in your windows PATH?
arg
>>101840122This is should not be needed as standalone Python and its libraries always have the path and venv based installations have their own local path.
>>101840096--cuda-malloc perhaps or ComfyUI.
>>101840135Okay, then i have no idea.Ask ChatGPT. I've solved many issues with it
>>101840149Use your own brain, I've solved many issues with it.
>>101840160lmao. Sure that's one way to do it. It's pretty fast with chatgpt tho
>>101840171touch grass fr fr no cap
>>101840171>Carat !g0ld/24keYCarat Gold 24 Key Looks like a bot or a sucking account.Sucking accounts are the worst.
>>101840096What version of CUDA do you have?I had that same error and had to update from 11.8 to 12.1 to get it to work and some people have mentioned they had to drop from 12.4 to 12.1 to get it to function.
>>101840148That goes under "COMMANDLINE_ARGs"?
>>101840220Yeah launch parameters.
>>101840213I'm not sure, honestly. How do I check? Would I need the CUDA Toolkit for that?>>101840231Didn't seem to do anything.
>>101840306Open CMD, type in nvcc --version.It'll tell you what version of CUDA you have.
>>101840365It says that "'nvcc' is not recognized as an internal or external command, operable program or batch file."
>>101834548God, those eyes. She's almost real, Anons.
>>101840365He's running a local python installation. >>101840397Just install ComfyUI portable.https://github.com/comfyanonymous/ComfyUI
>>101840397That's because you likely forgot to put a space after nvcc. There's a space between that and these lines --, I know it doesn't really show in there.
>>101840413I did.
>>101840397>>101840419Hmmm, wonder why >>101840122
>>101840096You could probably try manually creating the venv then installing torch + other requirements
I'm going to try installing Forge, then see if that helps with running reForge.
I think he is trolling.a) unless a prepackaged software is working it's a hardware issueb) wrong install
>>101840531How do I create a centaur in Pony?
>Carpathian forest and ancient castle in sunset, mist, dark atmosphere, abundance of surreal nature, abundance of nature variation, abundance of plant variation, surreal landscape background, abundance of landscape variation, abundance of surrealism, surrealism, minimalism, abundance of color variation, vibrant colors, art deco, low gravity, ukiyoe, no gravity, many geometric patternsIn AlbedoXL this would create silhoutte patterns I have posted before because its been brainwashed.CinematicRedmond thinks strictly in terms of subject.
>>101840605Most of the time including "centaur" is enoughhttps://litter.catbox.moe/4nofqq.png
>>101840630Of course the aspect is different...
>>101840641I don't understand how AlbedoXL can do this but when I prompt some texture with Albedo it's shit.
>>101836990I like both. Comfy for workflows involving more than one model. Forge for tinkering around with prompts and less stable workflows, especially with controlnet. After I learned about ctrl+mouse click to pan when drawing a mask in Comfy, I started to like it better for inpainting too.
>nigbo stops posting>thread quality improves tenfold in both generalsreally makes you think
>>101840640Thanks! I'll try wiht my prompt.
>>101840475I tried running Forge and got a "RuntimeError: Your device does not support the current version of Torch/CUDA!"
>>101840690You know it's about persistence. Debo posts images and keeps posting them, you never post any images whatsoever.And this amuses you. Well let me amuse you further.
>>101840730Faggot
>>101840746I know you and you know me, don't change your typing for this.
i miss schizo anon
>>101840419Have you actually installed CUDA at all?Then again doesn't matter.Just do this:First you check if you have any CUDA stuff installed and if you do, then uninstall all of it. After that you go to program files and program data to remove the leftover Nvidia folder.Then you update your graphics drivers.After that you download this:https://developer.nvidia.com/cuda-12-1-0-download-archiveIt asks you about CUDA Visual Studio Integration, so click the link in the text and install visual studio.Once it's done, select the "Desktop Development With C++" in the visual studio window and just install all of themThen continue with the CUDA setup, don't install any of the outdated graphics drivers etc.. that it suggests. After that download CUDNN from here and install ithttps://developer.nvidia.com/cudnn-downloadsRemember to choose just the 12.1 options, don't install the 11.8 that comes with it.Okay now you have the CUDA stuff sorted out.When that is done, you go to your venv folder, go to scripts and type CMD in the address bar, type "activate" in the cmd and then update pytorch by writingpip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121Now you're good to go and this fucking program should finally recognize your GPU.You can also add --cuda-stream to your COMMANDLINE_ARGS= However do NOT type --xformers in there, because that will fuck things up for you. Xformers is causing issues for this one.Screencap this as a reminder how to unfuck the "Cannot use torch" bullshit.I did this yesterday because of myriad of errors I was getting.Oh yeah one more thing. If you have an older GPU than the 3000 series, I have no idea at all whether this stuff is compatible or not.But I know it works for the 3000 series and higher.
>>101840800I'm sorry anon I cannot read text especially if it's that huge...
>>101840798Expressing self-love. Noted and posted forward to your own state's mental health clerk.
>>101840800>Have you actually installed CUDA at all?I honestly don't know. If Command Prompt didn't recognize it, then I believe I don't. But how do I make sure that I actually have any CUDA stuff installed?
>>101840847If you have NVIDIA gpu you have CUDA installed.
>>101840800>>101840810that looks like a bot response to you lel
>>101840862But how do I check for that?
>>101840864Yeah I think so too. There's been GPT replies on /v/ too at least. I don't go to /pol/ so I guess I am safe here.
>>101840800CUDA runtime is installed by the drivers and that's all you need, he's not building packages, stop leading him down a rabbit hole.
>>101840878Stop trolling, I don't care.
>>101839984Flux nailed the T. Rex hands. LOL
>>101840916I'm going to post my version soon... it's not animal. But it takes to render.
>>101840930*a while to render* distracted
>>101840948Wrong render.
Anyone can recommend some decent hard-surface sci-fi models? I need to generate some concepts for work. Can be stylized, but not anime.
>>101840847Just go to "Add or remove programs" and see if there's any Nvidia cuda stuff present you can uninstall.If there isn't then just proceed without worry, if there is then just uninstall it and then proceed.I do have a feeling you don't have anything in there if you don't remember installing this stuff.>>101840864Lol, poor bugger got confused by the information load.>>101840881Would be nice if it worked like that, but it doesn't.I have fought with this damn program and my system for so many times by now, that I know for certain this doesn't work like people say it should.>You don't have to do thing X, it'll do it on it's own, you just need to..Never worked.Only by doing it manually like that have I got things properly running.Any kind of automatic install with this either hasn't managed to get things done or has just ended up shitting the bed by doing some half assed job that works when it feels like it.For example I didn't have any problems for a good while until the latest Forge update that required the 12.1 CUDA, not that anyone even said it's necessary.And I'm not alone with this.There's a ton of people who have tried a million solutions and only going by the manual bit by bit route things have finally worked.
>>101841036CinematicRedmond is great for film and other stuff. Because it's not brainedead as others.
>>101841036What work? What concepts?
q
>>101841043Thanks, will check out.>>101841059I work in video game art. There are often design proposals I need to make, and images help a lot. In a AAA you'd give the task to a concept artist, but in indies you take what you can get.I need model to be able to generate decent hard-surface sci-shapes that will be used as inspiration for guns, armor, devices and so on basically. Style doesn't matter much. Can be photorealistic, can be a drawing.
>>101841174I have worked years in vfx but luckily I am retired and besides I was in effects.You need to study how to make the first images and then how to bolster your own work with them.It will take 6 months I guess.Install ComfyUI it's like Houdini (but in wrong orientation) and use Photoshop.
>>101841174You cannot possible skip your own line of work at any point. Otherwise you would be fired.
I updated my graphics drivers, which got Forge to work, but reForge still won't work. Upon cloning assets, it gives me an error that "'git' is not recognized as an internal or external command, operable program or batch file."
Our problems are solved!https://videocardz.com/newz/nvidia-geforce-rtx-4090d-with-48gb-and-rtx-4080-super-32gb-now-offered-in-china-for-cloud-computing
>>101841209Been doing that for the past few days. From what I can tell so far, a lot depends on what one wants to render and the checkpoints/lora, hence me asking for some model recs specific to sci-fi.>>101841243I'm a 3d artist primarily, and doing this shit because they aren't willing to hire a full-time concept artist. Doing this with full support of the management.
my fucking fan broke>boring photograph of a cluttered apartment room. in the middle of the room is a broken floor fanalso, I don't think AdaptiveGuider works for negatives because I have 'window' in my negative>>101841266>china:(
>>101841311Just concentrate on what is your speciality - i.e. they won't hire a general 3d artist for FX job and all that jazz. Are you a generalist?
>>101841311Ahh wait It's my English - they are too cheap - Just say you are not able to do the work they want.I can type lot of things, I was in London in a close pub and a graphic designer complained they want a perfectly rendered strawberry.I told him it's not possible unless you hire people who actually know what they are doing.
>>101841311Best option for you (let's forget the Finnish English here) - contact linkedin and get hired by a real company. I am not going to advertise these real companies. It is also different if you operate inside the US. I am from EU.
>>101841336Yep, doing a bit of everything, but mostly inanimate objects.>>101841371I've more or less told them that AI is definitely not a silver bullet, and most of the time I can't get it to render what we need. They seem to be understanding.But I definitely see that this can be a powerful tool for suggesting some design ideas that I can iterate on later with paint-overs and 3d model blockout. I seem to need some models that are closer to my needs. I understand that training requires quite a lot of computational power, but would be nice to train a model on my dataset.
>>101841451I hope you are not lev or some troll - I doesn't matter what you tell and what you are - you were hired to fill that one job and you will stay there until you feel comfortable to say - are they pushing you too much or do you enjoy the way they push you?Because sometimes push gives a purpose.
so on forge with flux, if the prompt is over 550 or so tokens, it'll show the "TypeError: 'NoneType' object is not iterable", soemtimes as an out of memory, but mostly as a "RuntimeError: stack expects each tensor to be equal size, but got [526, 4096] at entry 0 and [518, 4096] at entry 1" (the numbers are arbitrary depending on the length of the prompt). also a quick way to clear OOM in forge is to load a small checkpoint model after you gen and then reload the flux one
>>101841515>if the prompt is over 550 or so tokens
>>101841451You cannot be responsible for any company either my man.
>>101841531boomer prompting increases the token length sadly
not sure whats her problem is
>>101841563thats schizo prompting
Is there a vid 2 vid service I can use for free to test an animation I made?
>>101841586maybe a little
>>101841535I think was made by some chat geezer.
https://www.youtube.com/watch?v=Bc4guQ8MW7Y
I'm amazed that the strategy to contain the schizophrenia in this general has worked so well.
>>101841752I only care about my are and my music.
>>101841767I could swear I typed ART
>>101841752What is funny? You don't mean anything.
>>101841258Fixed it, made sure Path pointed to Git. Now everything's installed and it works. I apologize if I ended up wasting anons' time.
>>101840756>hoe so mad she became human
On this ship we are listing Mariya Takeuchi. https://www.youtube.com/watch?v=CgNIy3MTysM&list=PLhB53ktYgG5BF8QkPxF-Plr5ZWFKc7Fxp&index=1
>>101841933Fuck you I have Parkinsons.
succubus
containment thread.
>>101842183She's just like Catie Price :)https://www.dailymail.co.uk/tvshowbiz/katie-price/index.html
>>101842135>>101842183Did you ever ask yourself why the pastebin was made?
>>101842211pre-bogged
>>101842178Oh its the thread commentator who enjoys attention.
>>101842228Thank you for posting. I'm running out of ideas.
>>101842248succubi is always good filler content until the next idea comes along
>>101842263You are right. I love Mariya anyway... ;(
>>101842135Hips not wide enough, I'll keep my precious semen
>>101842349Itsudemo.
>>101842263Thank you, I am sorry I asked.
>>101842401I have posted this image earlier. It was a throwaway.
>>101842401Most of the retards are unable to deliver 1080p image.
>>101842453take the dinopill
I love Mariya Takeuchi.
>thread schizo on a report spree
>>101842400it seems hard to control proportions in flux. she comes out pretty slender regardless
>>101842768thread schizo
whats your favorite cfg value, for me its 6
don't look at the toes>>101842786I have a new workflow set up with dynamic thresholding + adaptive guider + flux encoder guidance. my cfg is 6 because I'm afraid to change anything
>>101842829>don't look at the toesBut you need attention and post it anyway thread schizo?
I wonder if the hands/toes are a product of nb4 or not
idk how I feel about cranking negative guidance
>my negative is blank in flux. not sure what to do.
>>101843080>>101842959>>101842829You might be getting too close to dangerous NSFW poses.
>>101843146I got a few see-through bras I can't post but I don't think flux does actual nsfw stuff
>>101843164>I don't think flux does actual nsfw stuffSomeone made a boob lora but it doesn't look much better than the deformed nips the model already can make https://civitai.com/models/639046/flux-booba
>>101843164That's why I said they are dangerous. They make FLUX turn into SD3.
>>101843164is there a way to use flux on a 3080 or do you really need more vram
I'm using it on a 3080. It works best with Forge now after the new update.
>>101843244yes, absolutely. I'm on a 4070 w/ an nf4 setup and I'm cooking at 2s/it. you can also run with fp8 instead of nf4 and gens will be somewhere between 5-15s/it from my experiencethere's more detailed setup info here:https://comfyanonymous.github.io/ComfyUI_examples/flux/or if you want to use forge instead:https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/981
baker-san
baker-san...
>>101843516>>101843516>>101843516
>>101843080for realistic stuff it doesn't work that well, you can make it better by using a low threshold on AdaptiveGuidance and by decreasing the max_shift value of 0.8 on ModelSamplingFlux, can you give me your workflow so that I can show you how it can be improved
