Thanks for all your encouraging words.
Now, let's get down to business
you didn't turn the sound into a distorted mess either
Audio tracks are from PS1 release. Then compressed to OGG at CBR 320kbps for quality reasons. Most FMVs are pretty short and the size difference between VBR at eg. target 128kbps is not big, so I decided I'll go for the
MaximumTM quality.
still a bit too blurry in some spot but that's huge progress for an automatic upscale..
I managed to get some sharper results in some cases, but it was a mixed bag. Sometimes the objects/characters became so sharp that they didn't really fit into the background (halo-like artifacts along contours). Like they didn't belong in the scene. I tried to avoid that whenever I could.
I focused mostly on animation stability, to reduce frame-to-frame wobbliness seen in some other FMV upscaling projects. Sometimes to achieve that I had to sacrifice a bit of sharpness. Because what looks better in still images, sometimes looks worse in motion. While my FMVs are not super-stable in motion, they are stable enough (at least for me
) considering the source videos.
I wonder if different upscale methods would give sharper results and that could help using animated mask to blend between different type of upscales.. that would require a lot of roto-masking but it could work maybe..
Yes, I actually thought about that. For example, some models give sharper looking characters but destroy the backgrounds by turning them into oversharpened mess. Other models produce great looking backgrounds, but the characters are blurry. So it would be the best to somehow auto-combine them together to achieve the Maximum
TM quality
Hope you don't mind I uploaded it to Youtube
Not at all.
Those are pretty sharp looking FMV's, what ESRGAN models are you using?
Me and MCINDUS made a custom model. I've done about 98% of FMV's upscaled to 4k, we could work together and achieve a quicker release.
I use a lot of different models, software and techniques on every single frame within a scene (and some FMVs have a lot of different scenes, sometimes I use different models between the scenes, for example, I used completely different methods for upscaling the waves from the intro FMV, different ESRGAN models for Seifer vs Squall fight etc.) Like I said before, it's a trial and error process to figure out what model/technique combination works best for that specific scene. And it's very time and PC resource consuming (I hope my GPU won't die on me, since it's working some crazy hours at 100% every day
)
I use models from here:
https://upscale.wiki/wiki/Model_DatabaseJPG PlusULTRA, DeJpeg Fatality PlusULTRA, ISO denoise v1 among others. Sometimes one of them, sometimes all of them. So one frame of FMV can undergo 3-5 different ESRGAN models. IIRC overall there are about 16k frames for Disc 01 alone. It takes a lot of time for sure. Sometimes I use them directly on source BINK PNGs, sometimes I upscale first and then denoise/deblock to achieve the best quality.
I'm still in the phase of prototyping/testing different models, different techniques and different combinations of them, so I'd only slow you down. I think you should continue with our work without me, since my mod won't be ready for the next 10 years or so (if ever
), given how I chose to approach the project. Your mod looks great too, and I can't wait to use it in the meantime.
