Upsampling Textures with AI
I did a little experiment with FPS Creator Pack #75 by upsampling the images with Gigapixel, which uses deep learning to upsample images and infer details that don't appear in the original pixels. The AI neural network does a pretty impressive job, generating results that are look better than a simple sharpen filter: I doubled the size of the textures to 1024x1024. Then I generated normal maps from the high-res images using AMD's TGA to DOT3 tool, and saved the normal maps with BC5 DDS compression. The diffuse textures were saved with BC7 DDS compression. The images below are using a 4x magnification to demonstrate the difference.
As you can see, the image that is upsampled with deep learning looks normal and the resized image looks like there is butter on the lens! It's hard to believe the above images came from a 256x128 section of an image.
The workflow was pretty tedious, as I had to convert images to TGA, then to uncompressed or BC5 DDS, and then to BC7 in Visual Studio. Each BC7 texture took maybe 5-10 minutes to compress! So while this set represents the optimum quality for 2019 game tech, and the format for assets we want to use in LE5, the workflow has a lot of room for improvement.
You can download the full package here:
- 1
6 Comments
Recommended Comments