Previous /sdg/ thread : >>101882520>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/ComfyUIForge: https://github.com/lllyasviel/stable-diffusion-webui-forgeSD.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
>mfw Resource news08/14/2024>Flux.1-Dev NF4 Quant v2: flux1-dev-bnb-nf4-v2.safetensorshttps://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/1079>ComfyUI-Lumina-mGPT-Wrapperhttps://github.com/Excidos/ComfyUI-Lumina-mGPT-Wrapper>bigdata-pw-Dataception: Dataset of datasetshttps://huggingface.co/datasets/bigdata-pw/Dataception>InstantX / FLUX.1-dev-Controlnet-Union-alphahttps://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Union-alpha>Judge Advances Copyright Lawsuit by Artists Against AI Art Generatorshttps://www.hollywoodreporter.com/business/business-news/artists-score-major-win-copyright-case-against-ai-art-generators-1235973601/>Integrating Saliency Ranking and Reinforcement Learning for Enhanced Object Detectionhttps://github.com/mbar0075/SaRLVision>ComfyUI nodes for ControlNext-SVD v2https://github.com/kijai/ComfyUI-ControlNeXt-SVD>ComfyUi UniAnimate: Taming Unified Video Diffusion Models for Consistent Human Image Animationhttps://github.com/Isi-dev/ComfyUI-UniAnimate-W>Stable Audio ControlNethttps://github.com/EmilianPostolache/stable-audio-controlnet>ComfyUI_NAIDGeneratorhttps://github.com/bedovyy/ComfyUI_NAIDGenerator08/13/2024>Kohya training to enable Flux training with 12GB VRAM GPUshttps://github.com/kohya-ss/sd-scripts/pull/1374/files>InstantX / FLUX.1-dev-Controlnet-Canny Updatedhttps://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Canny/tree/main>ClickAttention: Click Region Similarity Guided Interactive Segmentationhttps://github.com/hahamyt/ClickAttention>A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Modelshttps://github.com/taehong-moon/ee-diffusion>ComfyUI Dwpose TensorRThttps://github.com/yuvraj108c/ComfyUI-Dwpose-Tensorrt>SSL: A Self-similarity Loss for Improving Generative Image Super-resolutionhttps://github.com/ChrisDud0257/SSL>CogVideoX: Text-to-Video Diffusion Models with An Expert Transformerhttps://github.com/THUDM/CogVideo
>mfw Research news08/14/2024>Imagen 3https://arxiv.org/abs/2408.07009>Low-Bitwidth Floating Point Quantization for Efficient High-Quality Diffusion Modelshttps://arxiv.org/abs/2408.06995>Prompt-Based Segmentation at Multiple Resolutions and Lighting Conditions using Segment Anything Model 2https://arxiv.org/abs/2408.06970>SceneGPT: A Language Model for 3D Scene Understandinghttps://arxiv.org/abs/2408.06926>Dynamic and Compressive Adaptation of Transformers From Images to Videoshttps://arxiv.org/abs/2408.06840>Token Compensator: Altering Inference Cost of Vision Transformer without Re-Tuninghttps://arxiv.org/abs/2408.06798>Improving Synthetic Image Detection Towards Generalization: An Image Transformation Perspectivehttps://arxiv.org/abs/2408.06741>DiffLoRA: Generating Personalized Low-Rank Adaptation Weights with Diffusionhttps://arxiv.org/abs/2408.06740>DC3DO: Diffusion Classifier for 3D Objectshttps://arxiv.org/abs/2408.06693>Masked Image Modeling: A Surveyhttps://arxiv.org/abs/2408.06687>Hybrid SD: Edge-Cloud Collaborative Inference for Stable Diffusion Modelshttps://arxiv.org/abs/2408.06646>EditScribe: Non-Visual Image Editing with Natural Language Verification Loopshttps://arxiv.org/abs/2408.06632>Prompt Recovery for Image Generation Models: A Comparative Study of Discrete Optimizershttps://arxiv.org/abs/2408.06502>Synthetic Photography Detection: A Visual Guidance for Identifying Synthetic Images Created by AIhttps://arxiv.org/abs/2408.06398>Response Wide Shut: Surprising Observations in Basic Vision Language Model Capabilitieshttps://arxiv.org/abs/2408.06721>Breaking Class Barriers: Efficient Dataset Distillation via Inter-Class Feature Compensatorhttps://arxiv.org/abs/2408.06927>Do Vision-Language Foundational models show Robust Visual Perception?https://arxiv.org/abs/2408.06781>ViMo: Generating Motions from Casual Videoshttps://arxiv.org/abs/2408.06614
>mfw Fag Resource news08/13/2024>Doghya training to enable Flux training with 1GB VRAM GPUshttps://github.com/kohya-ss/sd-scripts/pull/1374/files>InstantX / FLUX.1-dev-Controlnet-Canny Updatedhttps://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Canny/tree/main>ClickAttention: Click Region Similarity Guided Interactive Segmentationhttps://github.com/hahamyt/ClickAttention>A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Modelshttps://github.com/taehong-moon/ee-diffusion>ComfyUI Dwpose TensorRThttps://github.com/yuvraj108c/ComfyUI-Dwpose-Tensorrt>SSL: A Self-similarity Loss for Improving Generative Image Super-resolutionhttps://github.com/ChrisDud0257/SSL>ZePo: Zero-Shot Portrait Stylization with Faster Samplinghttps://github.com/liujin112/ZePo>CogVideoX: Text-to-Video Diffusion Models with An Expert Transformerhttps://github.com/THUDM/CogVideo08/12/2024>Linux Foundation Welcomes the Open Model Initiative to Promote Openly Licensed AI Modelshttps://www.linuxfoundation.org/press/linux-foundation-welcomes-the-open-model-initiative-to-promote-openly-licensed-ai-models>PreciseControl: Enhancing Text-to-Image Diffusion Models with Fine-Grained Attribute Controlhttps://rishubhpar.github.io/PreciseControl.home/>ProxyCLIP: Proxy Attention Improves CLIP for Open-Vocabulary Segmentationhttps://github.com/mc-lan/ProxyCLIP>CSAM TURBO: Bonus oRthogonAl Token for Architecture Agnostic Textual Inversionhttps://github.com/jamesBaker361/tex_inv_plus>CumfyUI-RefUNet: nodes to use Reference UNetshttps://github.com/logtd/ComfyUI-RefUNet08/11/2024>Plush-for-ComfyUI: Mixture of Agents and other Agentic Systems Workflowshttps://github.com/glibsonoran/Plush-for-ComfyUI>CmfyUI-SEGAttention: Smoothed Energy Guidance Nodehttps://github.com/lgtd/ComfyUI-SEGAttention>SimpleTuner v0.9.8.1 - much-improved flux traininghttps://github.com/bghira/SimpleTuner/releases/tag/v0.9.8.1
First for debo is the thread schizo
>>101891466SD eyes
no one cares about the paper spam called "news"
>>101891521given how dead these threads are you wonder why he bothers posting them anymore
>>101891491who cares about eyes tho
2 years dudethink about that
>>101891625whom'st've am i talking to?
>nigbo
why is there a discord in op?
https://rentry.org/debo
>>101891580Catjak created /ldg/
>>101892040based
I like hlky's avatar, I think her lines are neat and the tattoos are better than ordinary tramp stamps
>>101892125Hi hlky
>>101892143no I'm not tho
>>101891440>CSAM TURBO
too annoyed to nap desu. i'll just have another coffee>>101892125thanks. she only has tattoos sometimes, they're temporary>>101892143hi
>>101892040please, call me ranfag
>>101892262ywcthe lines are nice and fine, I like that about it2 point for house big data
>>101892287He won
>>101892385Crazy, is this flux
>>101892396Yeah, it's very good, if only it didn't take 7 minutes to generate
>>101891440>https://github.com/lgtd/ComfyUI-SEGAttentionThis vanished off the face of the earth, but an alternative implementation exists with https://github.com/pamparamm/sd-perturbed-attention which I have been using. Yes, it slows down generation but it actually does look better.Additionally, I've been looking back at some older methods for reducing VRAM usage and came across Token Downsampling again, from February this year.https://github.com/ethansmith2000/comfy-todoThis did work with Flux but the math behind the downsampling is really confusing, I don't understand how you can get a ratio of 0.75 tokens merged or eliminated with a downscale factor of 4 and etc. Regardless, it is working better than ToMe in ComfyUI that is built in. I would make a patch to officially include it but all the code is open source with no license so fuck that.
>>101892419Is this done locally? 7 minutes is way too long.
>>101892385the background faces are impressive in their sharpness and quality, I think
>>101892425CumUI
>>101892435Yeah, it's done on a measly 3060 so that's why.>>101892438I agree, no inpainting or anything.
>>101892419What gear are you using? Is this with comfy or forge?
>>101892483>Yeah, it's done on a measly 3060 so that's why.Do you have catbox? I have a 4090 and want to recreate it.
>>101892359>got doxxed>incessantly bullied for his furfag/loli fetishes>goes to sleep/wakes up thinking about Debo/PWhe just can't stop winning
/sdg/ is healing
>model draws ear too big>inpaint it>ear gets 15% bigger...
Excuse the noob question, but where might i find low bit flux models?NF4 are on civit but fp8e4/fp8e5 do not seem to be there.Tried comfy site links but they crash.I have more than enough VRAM but for some reason i can't get it to work.
>>101892494It's an ultimate sd upcale with 4 tiles of 960x1216, euler 20 steps, probably take ~1 min on a 4090.
>>101892577you set fp8e4/fp8e5 with the weight-dtype dropdown, it's not a separate model like nf4
>>101892615>ultimate sd upcaleHow about tiled diffusion? Ultimate SD Upscale from my experience is extremely slow.
>>101892615where is little Margaret Thatcher from the previous gen?
i'm trying the bnb nf4 thing and it's using the modelsamplingflux node which has max_shift and base_shift params, wtf are these and is there any reason to care about them?
hello
>>101892425I thought token downsampling was more for secondary upscaling/detailing passes where granularity was less important. won't the downsampling impact adherence?
>>101892755howdy
>>101892724>wtf are theseno one knows
>>101892349Cozy
>>101892758You can do it for primary generation for the reduced resource usage, and I have been feeling my hardware is limiting what I can do with Flux resolution-wise. You are correct though as it's basically token merging improved.It's available also for A111 as well for those wondering.https://github.com/feffy380/sd-webui-token-downsampling
>>101892862ty
>****
It's been a while since I've used SD, what's a good model to save for the future that might get banned soon? I have SD1.4, SD1.5 pruned and wd-v1-2-full-ema-pruned, can't even remember what the last one is.
really reminds me of debo's multiday spergout when the basedbin was added to oponly with an older Debo who simply has no more power and has quite a bit of brain damage by now
do you ever look in a mirror and ask yourself what you are doing thread schizo?
>>101892489forge, nf4-v2 with t5 fp16>>101892658It's just the limit of the gpu, also xformers doesn't work currently so that doesn't help either.
>>101893048those aren't good models and they're still widely available loll
>>101893293Yeah that's why I'm asking anon.
>>101893048Now that you mentioned it, i wonder why no one made a Bstaber XL or Based64 Pony.
pathetic
probably drop what i have for flickr tomorrow with a note that the aim for the next version is 500m, should have like 250m+ by the morninghopefully big drop will finally start some recognition, diffusion1b is big but limited usage compared to flickrfor now i think maccies and x-files will cheer me up
like we need a blogger baka
>>101893337it really depends on what you want to do, vanilla sdxl and sdxl turbo, as well as flux and flux schnell are all really good to have for multi-step workflows as they're the best at prompt following and creativity, if you just want to 1girl some pony model + lora to taste will be your best bet, i like confetti and wai ani nsw for anime and pony realsim for realism, there are various other mixes probably worth tryingthis is done with juggernaut as an upscale pass
>>101893359Fucking hell I know I'm a lazy fag but when I fiddled a little bit with SD, all I needed to do was install the GUI and run the SD1.4 model for it to work.I'm downloading CyberRealistic_V5_FP32.safetensors now but it says checkpoint model for SD1.5. Does that mean I use it only together with SD1.5?
>>101893554no that's a full checkpoint, you don't need anything else
>>101893552>vanilla sdxl and sdxl turbo, as well as flux and flux schnellSo if I understand you correctly, those are general purpose models and the other stuff is I download a model with a specific woman already generated and create images from that?I didn't get very far probably a year ago but I remember generating the same woman with different clothes, leaving all settings as is except for the clothing keywords.
>>101893519we don't need nogens complaining or spamming their schizophrenic obsessions with other posters yet here you are
touchy hlky this is not /soc/
>it was hlky all along
>>101893607Ok thanks. I think I'll just download a few models and start to get into the topic again.It's been so long that I've never even seen half of AUTOMATIC1111s current options.
thread schizo
Anybody got a good pinup pose on simple background?
>>101893650this one goes into the hlky folder
first flux depth controlnet test... don't like that you have to use their sampler node
din jävla mamma
Horunge Bög
>schizophrenic
so /ldg/ only exists because anons wanted to get away from the thread schizo?
>>101894299right, and said thread schizo is currently spamming CP on there
>>101894299no, it started because pixart chang wanted a safespace and was co-opted as ranfag's blog to bitchpost about debo and pw
>>101894299>>101894322>>101894335/ldg/ is 4chan's version of a reddit hugspace>captcha: THIS
so its debo, gotcha
>>101894428I honestly can't imagine how sad that subhuman's life is though. I feel like one of these days he's going to disappear because he offed himself irl.
>>101894335Baking your cp on proxy debo?
maccies was nice except for the lunatic driver who got mad at me for stopping at crossing and almost ran over the guy crossing the road. wish i could live somewhere else desu>>101893913not sure how to feel about this
>>101894475nope, just a normal anon that thinks debo and ran are both massive faggots
it's no surprise that you are so lonely Debojust look at what you are doing
Mods please ban debo post on sight and he will leave /g/ for good
can someone link the 25% faster flux workflow that sets the CFG to 1 for the last half of the gen or whatever, i'm trying the dynamic thresholding workflow and it's def a bit better at prompt following but the gens are slow and things are coming out kinda overbaked(this is with the same settings as the "Hatsune Miku, 80's anime drawing style" example)
>>101894476many are not sure, but you're in good companyI am not some malicious person, I just like pictures
>>101894428it cannot be understated how important and influential I am
imagine having a pastebin about your past actions on fucking 4chan AND anons creating a new general just to get away from you LMAO
>>101894679I unironically hope you get a job working with immgen
>>101894720>joblol>nonslop ai genslmao
Mind the thread schizohttps://rentry.org/debo
>>101894679Is there a reason you are spamming cp on /ldg/?
>>101894720that'd be cool. suno woulda been my dream job but they left me on read
>>101894757can we cross reference the post timings? looks sus af
>>101894770why do they have such a long neck?
>>101894770openai have my dream job. i haven't applied for it though. i need more big data drops first desu
i wish flux was better at making girls thicc but the vibes are great
>the thread schizo containment general works
>>101894878not really, he's spamming cp on /ldg/
What's the distilled CFG for?No one knows?
>>101891401>train lora on flux output, no nudity except inpainted>the nipples start to disappear in some of the outputsam I being schizo
>>101894906I wish debo was hereHe could explain the situation and everything around it
>>101894861lol the flux gens always with the nails
>>101894838the official big data gang. I wonder how many dataset likes you need before openai is willing to look at a resume>>101894893>every poster is debo!!real schizo behavior>>101894906>distilled CFGall I know is what illy posted in his update write-up:>Flux-dev is a distilled model. It is recommended to set CFG=1 and then do not use negative prompts. Using “Distilled CFG Guidance” instead. The default value is 3.5. Note that if CFG=1, the UI of negative prompt will be greyed out.
>>101894861nice
>>101894965finally run out of ips?
>>101894965Thread schizo
>>101894965probably a lot more than i have atm desu
https://voca.ro/19HHTMCLdSYWtook a midday depression nap, and i am now refreshed and ready to .. discuss technology
>>101894835that's so they can take it deeper
>>101894988no. I'm omnipotent. I'm all knowing and all seeing. I'm every poster. I'm in your walls. I'm the last thing you see before you sleep and the first thing you see when you wake. I am your entire reason for existing. I am your God.
>>101894770yeah but whát exactly would you be doing there? surely not clicking 'gen' but coding for music generation?If you wanted to make it with the music one can still create an ai album and release it, or remaster that irl imhodesu
overcooked
>>101895089How does it feel to have catjak take everything from you?
>>101894679what exactly have you done besides rot in this thread all day every day for the past year?
>>101895051I read 'repression nap' and had to think of Britainpoor Britain, havin' a normal one
>>101895089>2 years of schizo posting
>>101895051>took a midday depression nap,nobody cares
nick fe ultimate cringe
>>101895125I was very well-suited for several of their swe openings. the intersection between music and AI is what would make working with them super cool but my actual work would be software. unless they have a "click gen all day" opening, then I'd prefer that
>>101895190The pastebin says otherwise
>>101895190I would take that position too, the click onewell there's bound to be competitors, one could pre-emptively market ones selves as suited for the future position of swe
>bimbo lipsgross
>>101891418>>InstantX / FLUX.1-dev-Controlnet-Union-alpha>https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Union-alphaNice
>>101895190>I was very well-suited for several of their swe openings.lmao sure you were bud... sure you were...
>>101895215pastebin?
>>101895215does it have archives of his posts where he admitted the whole reason he huffs comfy's farts is because he's because he can't into math or coding himself?
aight i'm gonna go on a walk in the forest but a casual suggestion, what if instead of every mentioning debo ever again y'all just let him beyou retards constantly talking about him are way more annoying than he could ever hope to be, ofc he's taking the piss w/ shit like this> no. I'm omnipotent. I'm all knowing and all seeing. I'm every poster. I'm in your walls. I'm the last thing you see before you sleep and the first thing you see when you wake. I am your entire reason for existing. I am your God.i would be too if people dedicated so much of their brains to me
>>101895345Holy newfag
>>101895282YesAlso the rentry in /ldg/ as well
>>101895282I can code fine, just not math in python. I couldn't write a stochastic solver if my life depended on it>>101895345you can't simply ask schizos to stop being mentally ill, but I appreciate you trying. enjoy the forest
>this is the place i want to be>a thread full of people that hate me
>>101895421>I can code fineyou literally can't lmao
>any new LoRas for Flux today?
https://youtu.be/LQhX8PbNUWI?si=UKT6cMduAWX1_IkV
>>101895421>I can code finepost github
>>101894965>Flux-dev is a distilled model. It is recommended to set CFG=1Thanks. That's a good starting point.
>>101895676you have to be joking
>>101895713This you?https://desuarchive.org/g/thread/96474271/#q96477123Please read the rentry/pastebin before engaging newfriends
>>101895275all those proxies just to spam report an insult on 4chan
>>101895888>he's wearing a different girl on his shirtlel
>>101895842no need to whine about rules (occasionally) being enforced.
>>101895842Oh he's doing it again? That's why we know he's the cp spammer
It's getting sad at this point
>>101895912don't make it too obvious
Hello again, anons! Totally passed out haha
>>101895974hello
>>101895907she was not happy desu
>>101896048Hey again, Debo! You were right hahaha that was way too early for me to be upThat gen is so cute! :]
>GIVE ME BACK MY PLUSHIE YA SLOOT
Flux's hand gen is great, the face and skin is weird
I know anons are angry about what happened to /sdg/ but please don't bother yelling at them.
>>101896134No one is yelling here, what are you talking about? /ldg/ brought it on themselves. This is a safe space now.
what is happeing today, this AM, this site was down, now github.
>>101896121Think I'll stick with SD1.5 until I read up more on how to get Flux to gen better esthetics.
she made me swap purple witch plushie for blue dino plushie
>>101896186Cute witch in the back haha! Oh yea I wanted to ask you, is that an easy way to interrogate hella images at once? It's so tedious the way i'm doing it LOL>>101896266Hahaha cute!!
>>101896235you should just jump right in and play with it and figure out as you goeven sdxl had models with great aesthetics superior to humu
>>101896299there* whoops haha still half asleepTime to get more coffee
>>101895751>no replyKek
>>101894587srsly does anyone have this, i remember it was posted a few threads ago
>>101894587>>101896450cant help you there, anon, no idea what you're talking about
>civitai down once again
>>101896299clip interrogate? what are you using? it's fairly simple to put together yourself with batch support if there isn't already and caching/reusing the text embeds if it's not doing that. the library i used to develop could do it fast desuclip interrogate isn't great anyway, the standard method depends on the content of the list, like it returns the most similar out of that list but the text may not actually be similar to the image. it's a bit better if you get the actual cos sim value and use thresholds but the appropriate threshold depends on which clip model and it's still hit and miss. this is why the stable diffusion safety checker doesn't work properlyyou'd be better off using something like florence-2-large for real images, ive liked the detailed_caption mode so far. for anime i think there are caption models based on danbooru tags
>>101896556flux is 2x slower if the CFG isn't set to 1, but there's "one weird trick" that makes it better at following prompts that works by setting the CFG to ~6 and the negative guidance to 10 (but still leaving it blank)some anon figured out that you can set the cfg to 1 for the later steps when it doesn't really change the output but it makes the overall generation faster because you're not getting the speed penalty of non-one cfg for some of the steps
>>101895751Debo sisters?
>>101896616>Something is going on behind the scenes today, gut says more consolidation.
>>101896615>>101896598do you really think this is how it works? how distorted is your mind?
>>101896630maybe do a search on one of the archive sites
>>101896630you want adaptiveguider
>>101892892Is that flux? because if so it looks really nice
>>101896617Yeah! Uhh something on huggingfaceIt works well enough, but it's super slow so it's kind of annoying and totally makes me not wanna do a huge thing with hella images to trainThanks so much! I think I have Florence already downloaded too :]
nice placement lel
>>101896730there's https://github.com/rom1504/clip-retrieval made by one of the laion guys, it's a bit overkill and not really designed for clip interrogation, the clip-inference part would let you embed images fast which you could then run interrogation onbut yeah better off using a caption model
>>101896852>caption modelWhich one is recommended?
>>101896852Thanks so much!! I'll give that a look :]
>>101896642Anon, we get it, you're a jelly Artist that can't into local. POST A GEN.>IT'S ALIVE!!
>>101896937What does that have to do with him lying?
>>101896307>humuHumu is just a great merge which yields excellent skin texture. SDXL has some great models as well but it's very fragile outside of specific parameters, reverting to that blurry plain skin look.Perhaps you can prompt for better skin texture. I'll try that on flux.
>>101896952what lie
>>101896952This isn't High School Anon, we don't give a @#%.. Post gens.. kek
>>101896957if you're on a111/forge, i found using a bit of dynamic threshholding brings out more details like skin on sdxl, like a cfg of around 10-12 and scaling down to whatever the model likesi kind of get it tho, i had one model i liked on 1.5 (i think it was yafm or azovyaphotoreal) and didnt bother switching to sdxl til i found something comparableyou cant really beat flux with its benefits tho, but i do expect checkpoints and other things based on it will start popping up soon
>>101896971Read the chain>>101896973We did and debo harassed everyone to /ldg/
>>101896875i've tested a few others recently, i was out for lunch using hf spaces on my phone so i can't remember the names other than moondream, the others were all too verbose and moondream made stuff up. i'd recommend florence. it supports other useful tasks like bounding boxes around objects and ocr, base is like 0.2b and large is 0.7b so it runs really fast>>101896926i'd stick with clip-L desu, or even use clip-B. H, g and bigG are just too slow for the image part. i did a lot of image embedding before for some project and i remember getting 13 image/sec with H when L was like 100+
>>101897010oh and you can use a bit of dynamic threshhold on flux as well and another extension called detail demoni know comfy has dynthresh nodes which are kind of similar (they seem less effective than forge's), and i'm sure there's a bazillion other nodes you can play withguess i just dont want to have to keep testing models and things when flux does 80-90% of what i want as of now so i'm good with just playing with the new hotness for now
>>101897031>florenceYou mean this one?https://huggingface.co/microsoft/Florence-2-largeIs there any workflow on how to properly use it?
>>101897014>We did and debo harassed everyone to /ldg/I've been here since the start and splitting the thread is dumb. Just have a Latent Diffusion General, encompassing all models etc. I dunno..
>>101897055this seems reasonable desu, but i came here after the threads had already split. just seems like discord drama that's unnecessary for 4chan
>>101897010>checkpoints and other things based on it will start popping up soonInteresting. Would be great if that was allowed. I'm not sure it is. Didn't read the licence. Would be great to have the hands of flux and the textures of humu.Prompting skin texture doesn't give me what I want. Nice composition with flux.
>>101897010>>101897099>checkpoints and other things based on it will start popping up soonhttps://civitai.com/models/645943/fluxunchained-artful-nsfw-capable-fluxd-tuned-model-by-socalguitarist
>>101897054yeah, or the base version, see pic related for base examples and above for large, the short caption is better in some cases with base but it misses details like text on detailed_caption, i'd say detailed_caption on large is the bestif you mean ui integration i'm not sure if there is yet, there's example code on the hf repo
>>101897099not the pro version or whatever of course, but the dev and schnell are apache license i believe>>101897106>Total training was about 5k images spread across SFW cinematic stills, art photography, LaION art-pop, about 1500ish explicit and artful nudes about 80% photography, 20% AI/illustrative nudes.seems like a good start
>>101897108Alright thanks anon
>>101897132Same guy who created the VisionXL models so I thought it's probably a good shout
>>101897163just wait >two more weeksand you'll see some serious shit, just based on how fast sdxl models came out>20+ gb per modelwelp time to free up some more space on my drive i guess
>>101897139here's what i mean about moondream, ive pasted the original image in because gradio hides the corner with the labelsome of the details are nice like 'vintage' and 'cobblestone street', not sure on turquoise vs green but if it's going to make stuff up like 'windows in the background' all the captions would need reviewing/editing. at least with florence everything seems to be accurateone of the other models i tested with this image said the one of the building doors was white, or that the car rims were black
>>101896616catbox please dearest anon
>>101897031I just found this thing, and it seems to work out pretty good! It loads up the text in 1 second haha>https://github.com/Hangover3832/ComfyUI-Hangover-NodesThanks for the help! I really appreciate it :]
>>101897448nta but there is something else to look out forhttps://huggingface.co/spaces/fancyfeast/joy-caption-pre-alphaIt's yet a little jank but it actually captions erotic scenes, hopefully that guy will really polish it. you can just grab the weits from the files section and run locally.
>>101897404Love this any chance of a catbox? Use any Lora's or anything?
>>101891401can you combine loras, eg angalina jolie punches brad pit?
i switched to kolors and flux based on /g/'s recomendations and XL is way betterthe fuck is wrong with you slutstoo much cock in your pussies?you gone crazy?need a lil foot rub to fix your brains?
/o/ retard here /aicg/ sent me here. I feel like an absolute boomer asking but how or where do I generate all those AI pictures that are everywhere on the internet, I'm guessing there are text to image generators but which ones are the most popular/accurate. I have seen some images on /pol/ of indans riding rockets made out of shit and stuff like that. How the fuck you generate something like that, please explain sorry for being a retard. Just checked the sticky and metastable looks like the most user friendly but I can´t see the requirements. I'm using a laptop with 8 GB ramn ryzen 5. I don´t need anything fancy just want to generate random images for fun. Any hope for me or should I go back to ooga booga wrenching
>>101897638You're better off using online services instead. Head over to this general >>101896094
>>101897638just use bing's image generator, it's for casuals and apparently decent enough
>>101897540Whoa, this one is super detailed! Unfortunately it took quite some time tho, and i'm too lazy to wait hella long hahaha!I might try to run it locally tho!Thanks, anon! :]>>101897581You can add both LoRAs and then proompt for the punching haha>>101897638Welcome to /sdg/, anon! I used to have 8gb and it worked just fineA bit of a wait tho hahaI would recommend using Comfy for your UI>https://rentry.org/comfyuithere are examples of how things work too :]>https://comfyanonymous.github.io/ComfyUI_examples/
>>101897703>hahapls just dont type it, you are are sounding like a creep
>>101897719haha
>>101895635What a disturbing picture, damn!
>>101897719had to do it, anon>>101897777isnt it? nice quads
>>101897677>>101897682stupid shir requires to log in with a microsoft account I think I will try >>101897703
>>101897719LOL get over it, anon!
>>101891401hey anons, does anyone know what model https://www.pixiv.net/users/8786609 (some nsfw warning) is using?peeps over at /h said it was novelai, but since when can novelAI pull such styles?
>>101897629>breath taking graphicskek, sonovabitch
>>101897879i think you can do like 3 for free per day without account, no? i never used itwith your specs, you wont be making but one image every 10+mins so not much difference
>>101897949>novelAII mean NAI does have one of the better anime models. Their latest one destroyed sdxl in terms of prompt comprehension and style. More than likely Flux will be able to compete once anime finetunes/loras begin to mature more
'haha' makes some ppl very upset
flux just gave up on realism for me lel
>>101897997Hope so, because image related is a cool style, love the whole 3d with 2d thingy, and haven't been able to mix a model that can do that at all, unless cheating with the 3d tag but that makes everything ugly. Is creating lora's on AI images still a no-no? Could I get away on training a lora on his gens?
>>101897997You sound a lot like a shill.
>>101898058I'm a Jew who hunts the souls of Muslims I'm a shilligami
New Thread!!!>>101897973>>101897973>>101897973
Filled!
>>101895998worrisome toes