Previous /ldg/ bread : >>101038464 >Beginner UI Fooocus: https://github.com/lllyasviel/fooocus EasyDiffusion: https://easydiffusion.github.io Metastable: https://metastable.studio >Advanced UI Automatic1111: https://github.com/automatic1111/stable-diffusion-webui StableSwamUI: https://github.com/Stability-AI/StableSwarmUI InvokeAI: https://github.com/invoke-ai/InvokeAI ComfyUI: https://github.com/comfyanonymous/ComfyUI >Auto1111 forks SD.Next: https://github.com/vladmandic/automatic Anapnoe UX: https://github.com/anapnoe/stable-diffusion-webui-ux >Pixart Sigma & Hunyuan DIT https://huggingface.co/spaces/PixArt-alpha/PixArt-Sigma https://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiT Comfy Nodes: https://github.com/city96/ComfyUI_ExtraModels >Use a VAE if your images look washed out https://rentry.org/sdvae >Models, LoRAs & training https://civitai.com https://huggingface.co https://aitracker.art https://github.com/Nerogar/OneTrainer https://github.com/derrian-distro/LoRA_Easy_Training_Scripts >Animation https://rentry.org/AnimAnon https://rentry.org/AnimAnon-AnimDiff https://rentry.org/AnimAnon-Deforum >Index of guides and other tools https://rentry.org/sdg-link https://rentry.org/rentrysd >View and submit GPU performance data https://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html https://docs.getgrist.com/3mjouqRSdkBY/sdperformance >Try online without registration txt2img: https://www.mage.space img2img: https://huggingface.co/spaces/huggingface/diffuse-the-rest sd3: https://huggingface.co/spaces/stabilityai/stable-diffusion-3-medium >Share image prompt info https://rentry.org/hdgcb https://catbox.moe >Related boards >>>/h/hdg >>>/e/edg >>>/d/ddg >>>/b/degen >>>/vt/vtai >>>/aco/sdg >>>/trash/sdg
>1.5 is the best for realism
Does anyone know what "DPM-SOLVER" and "SA-SOLVER" for PixArt are and where to get them?
>>101046667maybe >>101046686 ?
>>101046647Reality is an illusion.
>>101046757You're not wrong (and I'm not talking about living a simulation/matrix type bullshit)
>>101046720Nice! Cheers m8
>>101046788no problem
>>101046720wait a fucking minute... are you telling me I've been using the wrong ones?
>>101046667diffusion probabilistic model sampling vs stochastic adams solver. the second is faster but you can only use it if your name is adam
>>101046804Dunno. Most probably use extra nodes.
>>101046800damn it, I just wanted the scheduler. It won't work with my workflow.
ded
bred
I thought SAI died, and Stable Diffusion with it, but it seems what died were local diffusion models.
busy training
Any prompt help to get just shower glass door reflection? I tried silhouette and didn't work.
where is collage anon?this is fake bred
>>101046579
>>101047567try (stylized silhoutte trough shower glass)
>>101046579>BREADLY EVENT ANCHOR (Abstract Edition)img2img this img, and reply with the results!Recommended denoise: ~0.75
https://arxiv.org/abs/2406.11831>Benefiting from the inherent ability of the LLMs and our innovative designs, the prompt understanding performance of LI-DiT easily surpasses state-of-the-art open-source models as well as mainstream closed-source commercial models including Stable Diffusion 3, DALL-E 3, and Midjourney V6. The powerful LI-DiT-10B will be available after further optimization and security checks.HOLY SHIT ARE WE BACK??
>>101048806>10b image model + 7b LLMno one will be able to run that (even with a quantized LLM)
>>101048806the example images look incredible, hopefully they'll release something usable for vramlets too. can you offload image models to ram like with llms?
>>101048887you can offload everything with ram, that's the low-vram option on A1111 or comfyUi, it will just be really slow thoughThe advantage of transformers architecture is that it's resiliant to quantizations, maybe we can quantize DiT models, is there a paper about that already?
>>101048904>maybe we can quantize DiT models, is there a paper about that already?there's thishttps://arxiv.org/html/2405.19751v1I'm too much of a brainlet to see if it's good or not kek
>>101048806>We assign the text prompt of “a photo of {class}”But why. Sounds antiquated and takes me back to early days of dreambooth.. I bet their ass not everything they have in a dataset is photography, so why confuse the future tokenizer. I guess it's mitigated by mixing these captions with synthetically generated ones, and whatever's in that CC12M.>>101048887>the example images look incrediblePicrel looks on par with the rest desu
>>101048962I don't understand why they prompt for atmosphere instead of mood or something like that
>>101049027>they prompt for atmosphere instead of mood or something like thatNot sure what you mean. There's overlap between the two words, but "quiet and peaceful" or "terrifying" do count as moods I'd guess?
>>101049114>"quiet and peaceful" or "terrifying" do count as moods I'd guessFor sure. Perhaps I'm too used to 1.5 prompting. There's atmosphere and literal atmosphere which 1.5 mixes up all day everyday. With 1.5 it's better to describe mood, vibe or theme
how's it hangin anons?
>>101049138I'm still not sure how you differentiate mood from atmosphere. For me, atmosphere relates moreso to visual hints, and moods are well.. moods, emotions. Glimmers of sunshine and chromatic abberation might be atmosphere pieces, as opposed to more vague, emotional notions like peaceful, serene, or breathtaking.
>>101049484Very nice gen, don't mind if I snatch it for the next collage. I've been really impressed with some of the realistic gens lately, and those use to be my least liked back in the day. I'm still biased, but I'm starting to grow fond of the likes you just posted. Mind if I asked what model?I keep catching myself on biting the bait, in spite of reminding folks not to do the same. Good otherwise, one thing to worry about less since yesterday. How about you?
>>101049484training my 100th lora today. cool gen>>101049490yeah well clip seems like an autistic zoomer on twitter who takes everything literally, nuance is often lost
>>101049503Im the exact opposite! Ive always chased realism all the way from 1.5This is just PonyRealism2.1I hope no janny thinks this is too much for /ldg/
>>101049027I'm not fond of describing the atmosphere subjectively."foggy dark scene with occult signs and chromatic aberrations" ok"quiet and peaceful and snesual" wtf now you're just adding random effect shit with no control
>>101049558Very tasteful, you should be fine. I've been warned for less. Wish some anons understood that this is the beautiful way to go about posting and censoring risque gens.
>>101049590>>101049558lets keep it sfw hehe
>>101049617we have Malenia at homesorry Miyazaki
>>101049661
>>101049661lol, lmao even
>>101049661you should have to grind that "generate" button at least for a few hours, it's excellent 'design'
>>101049809>>101049810I agree!
Are there scripts to train loras or embeddings for Pixart?
>>101049810I like what you've got going too. I would only recommend picrel for the eyes/face. Maybe adetailer if you want to automate the process.
>>101049558Is this with any refiner or inpainting or just an out of the box gen?
>>101049948nope, no refiner, only adetailer.wouldve loved to fix the fingers, no patience today
>>101049902onetrainer has support for training lora but I had some issues with some settings that would usually work. it did train one lora. maybe there were fixes meanwhile, haven't re-checked.the upstream repos also has lora training code
>>101048855Can I run the LLM on one 3090 and the image model on my other 3090?
What's the best model for realism from your experience? Is there anything that is good at making things look like shakeycam/found footage?
>>101050440i'm guessing that it would be easily possible, it's possible on other current mode and often even also how some of these models were trained
>>101050464Can't say what's the best overall, but when it comes to Pony, I'm very impressed with: https://civitai.com/models/372465/pony-realismThe couple of realistic gens anon posted itt are good examples of it's potential quality.
>>101048855>(even with a quantized LLM)What the fuck do you even mean? Just run a cucked Q4 model on normal RAM, it's not like you need 400T/s to generate ONE image.
>>101051117I can't quite make out the details. What's in it?
>>101051258Peak /ldg/ humour.
>>101051235the 10b imagegen model will be too much for a 24gb vram card though
>cries in 8vram
>>101048741
>>101050582There's plenty good photo models but they all have a studio photoshoot, staged artistic look, probably because most of them are trained on stock photos and professional photoshoots. I want to achieve something that looks like it was shot with a bodycam/dashcam/amateur smartphone or other rather unprofessional type of photo. I want to do something like real looking cryptid footage or stuff like images of expeditions to another dimension/planet/fantasy world etc but captured in an amateurish way and showign mundane everyday scenes from these places. So far none of the models or loras I've tried really managed to achieve that look despite trying out all kinds of prompts/negative prompts.
>>101048741>>101051522>bodycam/dashcam/amateurHell yeah. This remind me of CCTV kinda gens, love those.Maybe you should look into these kinds of loras?https://civitai.com/search/models?sortBy=models_v9&query=cctv
>>101051522This one also looks promising:https://civitai.com/models/186174/dashcam-view-low-to-mid-quality
>>101051576>>101051618Nice, I can also try to make some Naked Gun intro style pics
>>101051658You can also try prompts and dutch angle, chromatic abberation and fisheye lens.
>>101051522I'm hoping the model will capture some of the amateur stuff. I've put a bunch of ghost hunting videos from YouTube into it. Ultimately the model should be responding to grainy, low light, grainy, poor quality, for images like this (screencap example). I also put a lot of urban exploration / liminal content into it.
>>101051937noice
>>101051937naisu
>>101051955If you have any example videos (ideally no watermarks) similar to this I'd happily add them into the dataset.
>>101051937right up my uncanny valley
Why is everyone in /sdg/ so on edge and angry all the time
>>101051937Very home videoesque.
how do you guys avoid sally sameface? all images I get,its the same girl, even with that detailer thingy its her again and again and again
>>101052131Impossible with the raped SDXL and SD 1.5 that are just merges of some autist's 10,000 steps of some girl's face.
>>101052145with 1.5 I had ton of embeddings I had trained of celebs and models, I just used to mash them together in prompt and got a unique face, with XL and Pony its just impossible to avoid her.looking for any and all suggestions!
>>101051937Reminds me of the first FEAR game
>>101052160It's unavoidable because the people who trained SDXL basically made the model memorize certain faces, there's no getting away from the AI Face. You can always try to put a random name in but the AI Face will always be there.
>>101051286how much would it need?
>>101052187damn, I was really hoping there was some way to avoid her.
>>101049977
>>101052187Are there any loras that let you manipulate facial features so you can choose from various facial shapes, change nose length/shape etc?
>>101052207SDXL (3.5b) asks for ~10gb of vram, so I guess that the 10b will ask for 28+gb of vram
>>101052251Im looking for those, if someone knows something pls link.since there are a ton of anime stuff, would they work?
>>101052131inpaint, except with a different prompt
>>101052304inpainted faces never really come out well
>>101052234
>>101049484>>101049558wonderful
>>101052332>>101052414thanks anon
>anon would rather shitpost than gen :/
>>101052131Use the old celeb name in the prompt trick. Maybe a wildcard of random names too.
>>101051937hype
I have a question bros.. Why do i keep getting this error despite my GPU having 6GBs ram (yes its 1660 but 1660 isn't that bad, right?) I also reduced the gpu ram usage to the minimum and trying to gen a 512x512 image, and still get this error. could it be related to something else other than the RAM?
>>1010531506vram is the bare minimum, and depending on what model you're using, they can be more demanding. With 6vram you probably want every little performance boost you can get. When it comes to models, you probably don't want to go beyond SD 1.5 or Pixart Sigma. Something like SDXL or Pony is probably out of the question. Guessing by the looks of your screenshot, you're probably using EasyDiffusion? I'm not sure how well optimized it is. You might want to try something like StableSwarm, Fooocus, Metastable, or https://github.com/lllyasviel/stable-diffusion-webui-forge
>>101053150Also if I'm not mistaken, something like 1660 probably misses out on architecture implemented with RTX cards, which doesn't help either.
>>101053275Yeah i'm using EasyDiffusion, was trying to use Pony. I will test with a SD1.5 model, and if the issue persists I will switch to Forge or something else, maybe easydiffusion is just not optimized..>>1010533141660 is pretty old desu, so it wouldnt be surprising that the gpu itself is the issue
>>101053440That would explain it. I'm sitting on 8vram of my rtx 2080 laptop, and something like Forge gave me the boost I needed for Pony/SDXL. Not sure if you can do the same, but who knows. The UI I mentioned should help with performance, because they're based on ComfyUI, which seems to be better optimized. If that's not enough, you should be able to gen on SD 1.5 models. Pixart Sigma is also efficient, but it's not supported by every UI yet, since it's a new thing. Either way, good luck. Machine learning unfortunately is VERY vram hungry.
>>101053150Run nvidia-smi to see if anything is using up VRAM.
>>101052529you make me want to try out pony again. the first time i saw your gens i almost thought it was midjourney fr
Realismanon, please teach me your ways.
>>101053150Why does it say 12.52 GiB already allocated if you only have 6? I've never seen the "already allocated" number be higher than what my card has.
>>101053918voodoo chakra magic, anon sacrificed his virginity to download extra vram
>>101053918I once allocated more RAM using my HDD to play some vidya that needed a much bigger RAM.. maybe it's that
>>101053912that poster already has a name
>>101054318what is it?
Finetune status?
>>101054337pedro
>>101054387pendejo puta madre
>>101054337Cardosanon/Troony
>>101054402no hablo espanyol
>>101054410its not comfytroon
>>101054410>Cardosanonif that is the case i will cut off my own penis and eat it
I regret asking, and I never will again.
>>101054447who the shit is cardosanon? dont bring sdg shit in here faggots
>>101054512Cardosanon is a founding member of this general and the discord
I've never really had a use for training my own model before so I know nothing about it, but in theory could I feed the thing an entire manga and make it replicate the mangaka's style exactly?
>>101054468ty for the reminder <3
>>101054518>discordlol
>>101054527You're welcome to join anytime
>>101046579>>101054527>Discord6wUwtcJsr2
>>101054598>>101054623no thanks
>>101054640Don't be retarded
>>101054660Rather be retarded than join that
>>101054468They never learn...
>>101054665So you are a retard then?
>>101054692if you say so
>>101054702You're the one who said it, actually, but yeah, guess I'm saying it too now.
Down the rabbit hole of deep intellecutal discourse.
there's a ldg discord btw
>>101054716its just that dumbass shockedmonkey and his troon army
>>101054525>in theory could I feed the thing an entire manga and make it replicate the mangaka's style exactly?"Exactly" depends on how skilled you are but you can get extremely close.
>>101054716invite?
>>101054716I wouldn't join a club that had the low standards of letting me in.
>>101052529>>101053906you should try it out its good!>>101054318you dont know me
>>101054824Any tips outside regular pony stuff?
>>101054959Photography terms seem to help,the detail lora is a must. Then its just following pony prompting -- that score_* thing.
>>101054518>and the discordGAY DETECTED! DANGER! GAY!
>>101055005Which lora is that? The old SDXL detail tweaker?
>>101055044thiscivitai. com/models/122359/detail-tweaker-xl?modelVersionId=135867
>>101055074cute
>>101054968these edits just show /ldg/ can't gen memes
>>101047880Futures looking bright localchads
>make SD3 tensorrt engine>try running it using workflows from https://github.com/comfyanonymous/ComfyUI_TensorRT/pull/30>get picrelwat do? it looks like a vae issue but i saved it from the SD3 Medium checkpoint
>>101054525the manga you're looking for is probably already contained inside a fine tune or a lora desu which one are you looking for specifically?
I can't believe Florence is almost uncensored.
>>101055286nice piclrel, anon
>>101055441you just need to turn up brightness on your monitor
>>101055441shit im retarded
>>101055631Maybe try a different sampler? I remember hearing some are not supposed to work with it.
>>101055631try euler
>>1010557210.o
>>101055649>>101055666still the same
>>101055734Maybe switch off sgm_uniform?
>>101055750
>>101055774Try connecting TensorRT Loader directly to KSampler, without thhe ModelSampingSD3 node?
>>101055935still the same, tried with euler,sgm_u and euler,normalweirdly enough SDXL models work properly with tensorrt giving speed improvements of around 80%
this is why I'm never going to use uncomfyUI
>>101056008I'm seeing a bug report from two days ago about similar behaviour. Try switching the amount of steps by 1, and maybe try updating the UI in case it was fixed?
>>101056008>>101056180except remeber to put that ModelSamplingSD3 node back while at it
>>101056180>>101056205the ui and the extension are the newest version, will try genning images with 1-30 steps one by one, with the ModelSamplingSD3 node
>>101056340oops didnt mean to upload pic
>comyUImore like SpaghettiUI
>>101056008is this an actual UI? lol this looks so autistic compared to Easy Diffusion's UI
>>101056180done genning 30 images with step counts of 1-30, each one is broken
>>101056457Because comfy is actually autistic and doesn't understand that nodes and settings that don't get updated often (if at all) should be able to be hidden. The basic node workflow should basically be a model selector, positive and negative prompt boxes and an output image, most of everything else is programmer garbage that should be tidied.
I'm installing StableSwarm in the background, and cmd is talking to me in three different languages. Is it trying to intimidate me?
>>101056457It is good tho.>>101056600I don't agree with every single node's design but in general this is the right way to do it.There are other UI that hide everything and serve users with only the most basic needs. Most definitely we're NOT just selecting models, positive/negative and output image, although if that's your thing you still can use the tinytera nodes or one of the alternatives.
>>101056423show me a different UI that does PixArt sigma, I'll switch.
>>101057075It's the right way in that these all need to be nodes. It's wrong in that it doesn't compartmentalize nodes and doesn't have global variables. And if you disagree you're another programmer with zero UX experience. I don't need to state that Comfy has a UX problem, because it does.
>>101057126SD.Next seems to also run it.
thank you for trying to help anons, guess its sd3 just being broken like always
>>101057142I am definitely not using this as a programmer.OK just do your better UI to show everyone your UX experience kek.
>>101057252>OK just do your better UI to show everyone your UX experience kek.Metastable and StableSwarm, to name examples that utilize comfy as backend.
>>101057252I don't know why autists who are fucking socially inept defend things they are fucking clueless about. ComfyUI universally is considered a shitty UX. It's not even close. I don't need to provide examples of good UXs, I only need to tell you that EVERYONE hates Comfy. And stop sucking the dick of things that suck just because that's all you can get. It's faggoty.
make a highly customizable workflow in a1111go ahead, I'll wait
>UI Warsb8888888888888
>>101057355Anon, you're describing visual programming. It's not customization, it's node programming. And as a node programmer, ComfyUI sucks compared to every other one. Stop defending shit just because it's a monopoly.
>>101057277You'd need MORE functions to deal with MORE complexity to be better. Only things that don't get in the way of this guideline are actually better.These other two UI just do less and are therefore simply worse. Or you can also see them as tools for a different audience that has different needs.>>101057301It's good. Feel free to make it better as per above guidelines or switch to the other UIs if you have different needs.
>>101057396>I use node-based ui's>I don't want to use node-based ui'sCurious
>>101057455Yes anon, I like Node UI's that are above the bare minimum. You know, like reusable node blocks? Or global variables?
Anon is right that the only reason comfy is used is because no one else has a "good" node based UI on the market. >invoke Lacks all the dumbass community nodes and addons hence not worth it even if it's objectively a better user experience
Which thread has higher gen quality? /ldg/ or /sdg/?
>>101057415No, I'm just commenting why Comfy will always be an autistic workflow that will never get traction. Mostly because ComfyAutist is completely resistant to any and all feedback so his product will remain mediocre out of spite and he'll eternally seethe about A1111 always having more market share despite have a literal monopoly on releases like SD 3.
>>101057503two different markets anon >casual consumers>autists, enthusiasts, "professionals"
>>101057521Wrong, just autists. Professionals will jump ship the second a professional, not amateur, node editor ships. ComfyUI is the Gimp of node editors, shit made for autists. Not enthusiasts. Not professionals. Autists.
>>101057540>Professionals will jump ship the second a professional, not amateur, node editor shipI don't disagree here. This is why they started the comfy org. He pulled the dev from stable swarm among a handful of other very talented devs. Good things are in the future for comfy.
>>101057503>will never get tractionit long did to the point where it shows up in random intel AI presentations and so onnow just go and make it better in a fork or make a better alternative from scratch if you are so convinced
>>101057562It's doomed because ComfyAutist is completely hostile to critique so he'll never change it because that is an admission of failure and apparently hostile to make good a good user experience. He has the "my way is right even when it's wrong" attitude and that's why he will always be Gimp and Blender of node editors (at least before Blender finally admitted maybe left click should be standard).
>>101057595Oh my it showed up in a presentation!?!?!?! That must mean it's great! ComfyAnon, your UI sucks and your org is going to fail because this is how you act every time your shitty UI is brought up. And you've barely changed it despite it being called out for being shit a year ago.
>>101057604>It's doomed because ComfyAutist is completely hostile to critiqueThis is precisely why he's bringing more people into the fold because he knows this is his main weakness (as he also previously stated).
>>101057636That doesn't help when you have a caustic "my way is right even when it's wrong" personality -- he's going to burn all those people out because he's going to fight to death about petty bullshit especially anything that he programmed no matter how much of a turd it might be.
>>101057685It's a possibility for sure, we'll see what time reveals.
Spaghetti is hard.
>>101057705I've been a professional long enough that I can spot the red flags. He's not a programmer that plays nice and absolutely hates real feedback that isn't sucking his dick.
>>101057632You brought up "traction" - it has very broad traction. You jump to "but my feels are more important"? Nah. Not particularly. Maybe if you make some improvement.
yes.... yes it is
>>101057725It has extremely limited traction and is struggling to gain momentum and you're too blind to see this. It has a bad UX and you have to have autism to use it. It will always be third place or lower.
>>101057770I'm guessing second place after the simple tool for "everyone".Until someone does a whole blender UI x IDE type work tool for AI + imagegen with far more changes to the python or w/e foundation too.
>>101056660Nice ritual
>>101057720Maybe try sucking his dick if you want some changes then retard
>>101057075>>101057252What's wrong with your gens anon?
>>101057953No I think he's a faggot who makes something as notorious as Gimp and Yandere Sim. I like to mock him and I'll delight in his continued misery of his self inflicted bullshit.
One more batch. One of the gens will be good.
>>101058304nice
> /aicg/ and /ldg/ unite!https://huggingface.co/facebook/multi-token-prediction
>>101058374>/aicg/Don't we have more in common with /lmg/?
Holy shit Florence is so incredibly good and it's fast. And it doesn't even censor penis.
Are there any easy-to-use webuis for Pixart Sigma, Lumina, and Hunyuan? It's pretty clear that stable diffusion is on its way out as one of the top opensource image generators, but I don't see anything other than comfyui for handling the other three mentioned on this page https://civitai.com/articles/5685/the-evolution-of-image-gen-with-sd3
>>101058538Thanks no gen, I'll check it out.
>>101058574https://huggingface.co/microsoft/Florence-2-large
>>101058538>it doesn't even censor penis.Which team do you play on?
>>101058538huh? I thought it was just a vlm
>>101058584I'm straight but I always test caption models on porn. Plus I'm making a model so I'm pretty equal opportunity when it comes to the dataset outside of the weird homo shit.
>>101058591Boy, you're gayer than two priests fucking an alter boy
>>101058596Yes, but that's good. It's fucking lightning, almost as fast as wdv3 and it's writing novels for captions. It's going to be the defacto captioning model.
>>101058608No, I'm gay as a daisy in may.
>>101058612Yeah I was retarded I thought you meant it was genning images. From my understanding (I haven't used it) that it is pretty damn good and punches well above it's wait given its parameter size.
>>101058629Its only problem is you really only get three captions: short, shorter and novel and there's no direction beyond that which sucks when you have a concept and you need to nudge it because it's a dumb AI. Especially for my Youtube video snapshots, Llava does really well being told it's a video and the title of the video.
>>101058629>waitweight
>>101058563SD.Next should handle the matter. Stable Swarm uses comfy backend, so it might be possible there. Metastable SoonTM.
Dare I say, is that a face?>1boy wearing jeans and a t-shirt
>>101058762What am I looking at? Other than a 1boy wearing jeans and a t-shirt.
>>101058769A 1.3B Pixart Sigma model being trained from scratch
>>101058779Why 256x256 then. Shouldn't you benchmark it at 512 at least?
>>101058790Because it's faster to train at 256 then 512 then 1024 then 2K than to start to train at 512. Almost all models are trained this way because if you give it too much information it takes way longer to figure things out.
hacked
>>101058810That's the cover image when I put it on Civit.
>>101058803Oh, that's very interesting. I heard about them being trained in progressive resolutions, but didn't think much of it in terms later finetuning by us. Back when I tried training loras for 1.5, I remember seeing better results training at 256 indeed. I guess there's method to this madness afterall.>>101058837Could get away with being a music album cover.
>>101058857I mean this isn't finetuning. This is a base model that was established from ground zero using my dataset.
>>101058867>>101058857This is the same prompt from yesterday.
>>101058762>>101058885Hype
>>101058885Slow and steady she goes.>>101058867I'm mildly confused. So.. you're not finetuning a 1.3B Pixart, but instead using the same dataset, method, and trying to replicate the results?
>>101058940So the original Sigma model's parameters is 0.6B, there's obviously limitations to that size. Using their research and methodology I've started a new model with expanded parameters (1.35B) which should be better. I've put together my own dataset of millions of images and I'm currently training it on two 4090s. The reason why I did 1.3B because it hits a batch size that's acceptable and should be doable even at 1024px. Besides having more parameters, my dataset is uncensored and has a mix of wikiart, datacomp (a new 1B llava 1.5 tagged dataset like LAION), e621, gelbooru, flickr, duckduckgo search images, and stills from youtube. Really whatever I feel like, there's a ton of video game content through ages, art, photography, etc. I'm trying to make it more than just a "uncensored" porn model.
>>101059013Coolness !!! Let us know how it goes!
>>101059013Oh, so you're that absolute legend I've been seeing around lately. Think I got confused because I didn't realise the parameter difference. Thank you for clearing that up. Once again, godspeed to your GPUs. Looking forward to the progress, however smol and incremental.
>>101058837release the model under faipl>https://freedevproject.org/faipl-1.0/
>>101059160Pixart is copyleft
>>101059176the license is too, check it out no joke
>>101059176>Pixart is copyleftas all things should be
>>101056340>staticHave you tried dynamic? You are using the right dimensions for the compiled engine but maybe there is a bug there.
>>101048806>and security checks.it's over
>>101059682>>and security checks.>it's overThose Chicom imperial state mothafuggers take security very seriously.
>>101058762Hype!!!
>>101058401yes. /sdg/ is exactly like /aicg/ as well.
sdg won
Anon sleep
>>101060806thief
>>101060594Isn't it time to merge back with the old thread?
>>101061030If they accept alternate, open models and SD corporate looks like they can avoid backruptcy.
>>101061030Why?
>>101060594not as much trolling as previous days hence slower thread >>101060612whats your definition of winning?
>>101061098Wasn't splitting the thread a NovelAI psyop to make local weaker? You know, the whole “divide and conquer” thing.
>>101061104popularity, relevance, skill
I like Sigma and I like SD models. They have their strengths and weaknesses. I don't like something that requires some remote resources that can change prices and terms on a whim. Having models that you possess and can run on consumer hardware is key. Don't get painted into a corner. Don't be dependent on a single supplier.
>>101061030Someone tried to have a joint thread but then someone else double baked an SDG so nothing really changed
>>101060866
>>101061030you left so you can't come back
anons are just in it for the drama but who can blame them. we all fall for that trap
Results are in for my first ever realistic skin LoRA for Hunyuan. Needless to say, the results are stunning. This is just with 100 quickly picked images where there is simple requirements: 1girl, no bokeh, mostly portrait, Asian. Unfortunately I was only able to find IG fashion thots, but as expected it has learned to make better images (that is, it learned how to make realistic skin). The settings were the same as they provided in their LoRA training script. As I said before, I think uncucking this model will be a piece of cake. Now, the LoRA is not perfect but it's a good sign of what is possible.
>>101061502
>>101061502anon..
>>101061502>Asian
>>101061513that ones real nice
>>101055074thank you for the amazing gens and tips, really great stuff.
Tbh it's time to merge these threads into one againNow that SAI is on its deathbed we can merge both threads to one called /ldg/Comfy leaving was the last nail in the coffin
>>101061680you guys lost and just need to come home already but you're too prideful
>>101061680i think anon is too attached to sdg, they wont change no matter what
>>101061745true but you have to hand it to them, they held out for a long time
>101061745>101061772Samefag
>this is bait
>>101061513Keep in mind I did not use just high quality pics for this LoRA, some of them are low quality, so it is copying the grain that it saw in some of my pics. For what it is, and how quickly I trained, this is not a bad result. Although a full finetune is still needed to fix details of faces in backgrounds.
>>101061853I'm going to have nightmares now.
>>101061844
nightmare nightmare nightmare
Fresh (for when anon wakes up)>>101061929>>101061929>>101061929
>>101061098Their mad that it's more difficult to run psyops in two threads as opposed to one
>318/108:(
