@majidzendaniyan lower the denoising strength
@majidzendaniyan lower the denoising strength
Θdisable Restore Faces
MIn this case, it does not give me quality with these details, Something that makes me excited and satisfied. There are unique details of hair and skin texture
Fthis isn't THAT specifically important. More importantly is that python 3.11.x isn't used. I had kernel issues doing AI using 3.10.6. A couple people helping me insisted that i used that to solve all my problems, so i downgraded from 3.10.9 to 6 and it actually started causing me hard system crashes. I suspect it was to do with incompatibilities between windows 11 on alderlake, and python but i never explored those suspicions. what i know is that the hard kernel problems go away when i changed python versions to a newer 3.10.9. i may upgrade to whichver the new 3.10.x is soon.
Θmine wouldnt run if python wasnt exactly 3.10.6
Fmine would run, but slowly, and then crash after a few staggery generations. i have alderlake and 4080 on newest OS. so we might have different needs. It may be a good suggestion for some people though right?
LIs there a way I can have SD start with specific settings already entered for the inpainting section of img2img? I pretty much always use the same prompt, sampler, CFG scale, denoise strength resolution, etc.
LAlso, is there a way I can start SD without internet?
Nstable diffusion is making my face wide, does anyone know how to make it squeeze?
Nnvm let me try randomizing seed
Pif you are running auto locally you aren't using the internet. your machine is pretty much talking to itself.
Fui-config.json in automatic1111 root directory allows the change. i don't think it's a specific change to the inpainting tab though, but all of img2img default value rather.
Fi think it uses what is embededd in model
Fyes please redownload model file
Fi think now it works . they made updates
Fam trying to train with kohya script and i i can only make it get this far. is there any help to advise here? it'll stay at this step all night and do nothing if i leave it.
total optimization steps / 学習ステップ数: 4350steps: 0%| | 0/4350 [00:00<?, ?it/s]epoch 1/15Fnot 8bit
Fdo you get any error?
Flooks like 0 steps
Fnone. it'll just stay there
Fi've had errors whiel i was building the data set and i learned to structure those folders right and got through it. this is a brick wall though. an error would aid. haha. trying adamw now
Fi dont know them but runpod pretty cheap
Fyeah same thing again. it fires up really fast and efficient. detects my image set. begins the training and just halts up. another thing i've noticed is if it's there and i press train in the gui again, it'll run again and get to that point. /shrug
Fvery interesting
Fdid you try with torch 1.13?
Fjust the default 1.12.1 with the cudnn fixes since i have a 4080. hmm will try 1.13
Fi tried the experimental 2 but ran into some startup issues
Nwhere to download pix2pix?
MWhich one should I download?
MI had downloaded the first one in the past
Nbig one if your gpu can fit
Fgoogle
Fpruned what
Npruned pix2pix
Nor the 7.7GB one
Fi don't use pix2pix, and if you check my picture you would see the model i was using
LIt will not start if I disable my Wi-Fi.
LThank you.
Fi did it! i got 1.13.1+cu117 to work! also copied the cudnn dlls into the torch/lib folder for good measure. it didn't work again, but there was a message that showed up this time! Right at the start of training before the rest of the steps like image set detection occur.
[WinError 127] The specified procedure could not be foundWARNING:root:WARNING: [WinError 127] The specified procedure could not be foundNeed to compile C++ extensions to get sparse attention suport. Please run python setup.py build developprepare tokenizerFbut working now?
Fnope!
i dont know what else to do . i think i broke it by trying to install new xformers. might just keep experimenting in the webui-extension but i can't seem to get styles to not look over trained there.
i dont know what else to do . i think i broke it by trying to install new xformers. might just keep experimenting in the webui-extension but i can't seem to get styles to not look over trained there.
Fno indication of what the problem is is really quite frustrating. not sure why the script won't log anything or error out or what
Fwhen comparing checkpoints (steps, methods, etc), what parameters do you guys use on XYZ? only seed and checkpoint?
Fi do batches of 4 so each cell gets 4 seeds anyways. checkpoint and cfg are ones i'll try. sometimes prompt s/r. you can try different prompts with it or different lora weights. the :1> can s/r to :0.5>, :1.5>, etc. more a lora thing
Fwhat cfgs you try? and why?
Fyou can try cfg
Fand perhaps pre selected seeds
Fadded some more prompts : https://github.com/FurkanGozukara/Stable-Diffusion/blob/main/Midjourney-Level-Realism.md
GitHub
Best Stable Diffusion and AI Tutorials, Guides, News, Tips and Tricks - Stable-Diffusion/Midjourney-Level-Realism.md at main · FurkanGozukara/Stable-Diffusion