3 ms·
It looks impressive, but what exactly is the machine learning doing on the original to produce the result? And wouldn't it be possible to simply take the origi
by zelon88 6y ago
It looks impressive, but what exactly is the machine learning doing on the original to produce the result?
And wouldn't it be possible to simply take the original minecraft map as a height map and texture map and then regenerate a new world with the original world data and more advanced post processing? You could interpolate and randomize more detail into the scene than you started with.
- erikpukinskis 6y agoIt’s not really adding any meaningful detail per se. where there’s a grass block it’s just rendering grass. All it is doing is projecting a stable image of “grass” (taken from a labeled image database) in that voxel. Not to minimize the awesomeness of that... doing it stably in 3D while moving the camera is the point of this paper, and is amazing. But it’s not really adding detail beyond “these are the kinds of pixels that grass has and the AI figured out we can put them in this arrangement without making things jumpy”
- debacle 6y agoIt also seems like your brain is doing most of the work here: https://nvlabs.github.io/GANcraft/images/vox2img.png https://nvlabs.github.io/GANcraft/images/vox2img.png The renderer seems to be adding some resolution, smoothing, and mipmapping. Shaders can do the same thing, and in real time.
- Bjartr 6y agoModern shaders[1] do a lot, but you'll never mistake them for anything but minecraft. They don't quite get to where this paper is demonstrating. [1] https://www.rockpapershotgun.com/best-minecraft-shaders https://www.rockpapershotgun.com/best-minecraft-shaders
- lainga 6y agoIt doesn't look like it can do structures, either.