How many steps do you guys do for supir?
How many steps do you guys do for supir?
SDoes anyone else ever run into an issue where upon generating an image the facial similarity is pretty good but after using adetailer its not even close? And if so how did you fix it?
F50
Fdo you set custom prompt?
Fto adetailer
Syes, ive tried a few things but the one i would expect to work (since again, its pretty accurate upon ggenerating the image) is ohwx woman
Fif you are using SDXL are you setting the resolution to 1024x1024
Sim using a custom resolution of 1024x1536 however ive had success with it with other dreambooth models ive made. and ive set the adetailer resolution to match at 1024x1536, yes
AAnything for linear stage 2?
Fdoes it make a difference if you set the adetailer resolution to 1024x1024, i've read that departing from the default res causing image degradation
SI trained the DB model on 1024x1536 so arguably 1024x1024 would be departing from the default res, but ill do some testing now on that resolution and edit this with results
edit: same result, adetailer comes back looking nothing like the initial results, much further from what the DB model was trained on
edit: same result, adetailer comes back looking nothing like the initial results, much further from what the DB model was trained on
Fcan i see the difference?
Fwhy train so high res?
Fare you generating the image at 1024x1536 and using adetailer at 1024x1536 too?
SI want to produce images at 1024x1536 so best to train for the resolution im producing. and yes i tried that and it actually worked out well, however i noticed that separate CFG scale was checked and although it was set to the same as the original generation i wonder if that was causing the error, experimenting with 1024x1024 for adetailer now
SBased on new tests, it seems to have been a bug with CFG scale even though it was set to the same amount as the initial generation
Fwhat denoising strength are you using
Si jump around from 0.4 to 0.6 but usually stay on the lower side
F0.6 is definitely going to change it a lot
Feven 0.4
Fset it down to like 0.25 and see how it looks
AWhat's the best ai tool for colorization of black and white photos? Does anyone know
SMaybe someone has a better recommendation, but if you use canny with controlnet and describe what you want, should do a good job adding color
FYouTubeHow Do?
This video provides a guide for colorizing and restoring old images using Unsampling and ControlNets in ComfyUI with Stable Diffusion. This is a follow-up of my video on "Reimagining" images and uses the same workflow with a few tweaks. For a more in-depth look at how that workflow works, you can watch that video here: https://www.youtube.com/wa...
FYouTubeAlbert Bozesan
To try everything Brilliant has to offer—free—for a full 30 days, visit http://brilliant.org/AlbertBozesan/ . The first 200 of you will get 20% off Brilliant’s annual premium subscription.
Links:
How to Install and Use Stable Diffusion (June 2023) - Basic Tutorial
https://youtu.be/nBpD-RbglPw
ControlNet: https://github.com/Mikubill/sd-webui-co...
Links:
How to Install and Use Stable Diffusion (June 2023) - Basic Tutorial
https://youtu.be/nBpD-RbglPw
ControlNet: https://github.com/Mikubill/sd-webui-co...
Fthere's some youtube videos about it (using controlnet)
DHi, please check out this workflow using CCSR as the first step, and then SUPIR as the second step to achieve superior results. Please see if such a workflow can be integrated into your gradio - https://www.reddit.com/r/StableDiffusion/comments/1b50sp0/ccsr_vs_supir_upscale_comparison_portrait/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button
Reddit
Explore this post and more from the StableDiffusion community
DExcerpt: "Conclusion
CCSR = high fidelity, but low quality (no fine details, washed out, softens image)
SUPIR = low fidelity (hallucinates too much), but very high quality (reintroduce fine details/texture)
CCSR ➜ SUPIR combo is simply mind blowing as you can see in example k, l, m. This combo gave the best fidelity and quality balance. CCSR is able to reconstruct as faithfully as possible even a destroyed jpg while SUPIR can fill in all the lost details. Prompting is not necessary but recommended for further accuracy (or to sway specific direction.) If I do not care about fidelity, then SUPIR is much better than CCSR."
CCSR = high fidelity, but low quality (no fine details, washed out, softens image)
SUPIR = low fidelity (hallucinates too much), but very high quality (reintroduce fine details/texture)
CCSR ➜ SUPIR combo is simply mind blowing as you can see in example k, l, m. This combo gave the best fidelity and quality balance. CCSR is able to reconstruct as faithfully as possible even a destroyed jpg while SUPIR can fill in all the lost details. Prompting is not necessary but recommended for further accuracy (or to sway specific direction.) If I do not care about fidelity, then SUPIR is much better than CCSR."
Fye looking pretty nice
PHello! i have a problem with SD img2img batch speed. I launch sd 1.5 Img2img with controlnet tile 200 images. at first it is super fast but the longer it goes the slower it becomes... the render itself is same speed but swithcing from 1 img to next is beocming slower over time. What can be the reason?
Pi got 4090 and SD installed on fast SSD
Fi think i can make a standalone installer for CCSR so that can be used to generate initial image sounding good?
Fvery possibly a bug
DYeah that would be great to try out
Fin a1111
Pyes
Fi would report this. i think there can be some memory leak
Fright now i am trying to make kaggle notebook of SUPIR
DNice
Fok it uses too much ram
Fso crashes notebook
Pi read this yesterday! the things move fast... haha
Fi think i found a good way to improve model loading of SUPIR
Fit was using 30 gb ram
Fnope
Vsorry it was a clickbait video..
Fyep
FGitHub
Currently create_SUPIR_model inside util.py loads everything into RAM first with full precision Thus it is using more than 29 GB RAM I think model half param can be given to it as well Or perhaps s...
Pi think you can use fp8
Pits use same stable diffusion engine that a1111
Pi am busy with work jobs i cant try it yet
