This is an ESRGAN model trained specifically for old video game textures.
While it specializes in GameCube and Wii textures, it can be used for other textures from that period, such as PlayStation 2, Xbox, and PC games.
Old textures often have artifacts that pose challenges for upscalers, particularly the Block Compression (DXT algorithm) still used in many textures today. HDcube 4 effectively removes block compression, color palette errors, dithering, color reduction up to 8-bit color depth, and 1-bit alpha depth.
HDcube4 is currently only available to supporters.
Comparison:
HDCube4-A
Color ⭐⭐⭐⭐⭐
Structure ⭐⭐⭐⭐⭐
Sharpness ⭐⭐⭐⭐☆
HDCube4-B
Color ⭐⭐⭐⭐⭐
Structure ⭐⭐⭐⭐⭐
Sharpness ⭐⭐⭐⭐⭐
HDCube3 500k
Color ⭐⭐⭐☆☆
Structure ⭐⭐⭐☆☆
Sharpness ⭐⭐⭐⭐☆
HDCube2 250k
Color ⭐⭐⭐☆☆
Structure ⭐⭐⭐☆☆
Sharpness ⭐⭐⭐⭐⭐
HDCube2 250k Soft
Color ⭐⭐⭐☆☆
Structure⭐⭐⭐☆☆
Sharpness ⭐⭐⭐☆☆
HDCube 200k
Color ⭐⭐☆☆☆
Structure ⭐⭐⭐☆☆
Sharpness ⭐⭐⭐⭐☆
HDCube 110k
Color ⭐⭐☆☆☆
Structure ⭐⭐☆☆☆
Sharpness ⭐⭐⭐☆☆
Usage:
chaiNNer if you use chaiNNer you should upscale the alpha channel separately.
Nice, this looks pretty good. I will have to try this and compare against something like BSRGAN which is admittedly what I use for most general purpose textures.
I had a chance to test it. When it works, it's a beautiful thing.
Sometimes though, it adds this graininess or sometimes even these lines to the image, not sure if I captured it very well: https://imgur.com/a/NQuLnVu . It can look nice on some images but others isn't as desirable.
I'll do more testing in the coming days. Certainly an awesome model to use in many situations. Nice work! You should post it to https://upscale.wiki/wiki/Model_Database
Yes, that's one of the things I'd like to improve.
Maybe I'll make a version that creates a softer image, so you can interpolate between the two and have more control.
but you can also mix it with any other model based on 4x_RRDB_PSNR
It looks like in a few cases it's turning dithering into noise. It might be worth applying dithering to some of the images in the training set so it knows not to amplify it.
Thanks for the feedback, I will consider dithering but I think I need to lower the gan_weight a bit.
I have improved my dataset a bit in the last days and the new version is even better.
here is a small comparison
I had to try it out on something (sorry PC not GC) that I'm working on.
[256x256 source, waifu2x, this model]
This was already a very smooth texture that fairs well under waifu, but what I like is how when this model picks up on sharp details on the nose and eyes. Highlights become very pronounced.
[color=#ffffff]https://imgur.com/Rbv8DqN[/color]