37
NVIDIA's Neural Texture Compression Cuts VRAM Use From 6.5 GB to 970 MB
(www.techpowerup.com)
This is actually a good idea.
Latent encoding/decoding is basically extreme, "good enough" image compression. It's a pretty old ML technique, and the models to do it can be small (eg fast-ish). A cross-platform implementation is doable.
If I were a game dev, I wouldn't overuse it, but it seems like a great way to store huge, somewhat less frequently-used textures in VRAM.
...But one thing I'm wondering is what they decode to? If its a huge bitmap (instead of a compressed texture format), that seems suboptimal. And I don't think it's decoded per pixel like BCx compression, unless I'm misremembering how it works. That granular access is kind of the point of texture compression.