As of right now the best models for CGI content within VEAI are Gaia CG and Artemis Alias & Moire. Gaia CG is good at maintaining image quality while upscaling and Artemis Alias & Moire is better at improving some aspects of CGI content. But both touch up on the entire image where it may not be necessary.
I’d like to see models similar to the specialized models trained by users of ESRGAN. These very specific types of models focus on one singular flaw and tries to improve said flaws and only the flaws while leaving the rest of the image completely untouched. Of course these models already exist if I’m asking for them here but ESRGAN is quite slow at ~4/seconds per image at 1080p.
These models are all 1x models and aren’t trained to upscale at all. Simply find the flaws and mend them.
Two ESRGAN models which show near perfection at their intended purpose are SS Anti Aliasing which was trained on Super Sampled video game renders and could perfectly make any CGI image look like it was rendered natively with SSAA.
And Bandage-Smooth is capable of not touching anything that’s not color banding. Typically very good at slightly compressed content that destroy color grading with bad banding and artifacts.