Hello everyone. I am Dr. Furkan Gözükara. PhD Computer Engineer. SECourses is a dedicated YouTube channel for the following topics : Tech, AI, News, Science, Robotics, Singularity, ComfyUI, SwarmUI, ML, Artificial Intelligence, Humanoid Robots, Wan 2.2, FLUX, Krea, Qwen Image, VLMs, Stable Diffusion
I have done extensive testing. I tested with your original implementation first then merging into my own gradio exactly same. Input image raw 256px Output final image Restored face output
Does 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?
im 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
I 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
I 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
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...
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.
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."