Realtime 3D rendering engine driven by a diffusion model
category: code [glöplog]
rloaderro: if you want image generator as versatile as Midjourney, they you need billions of params..but maybe you can narrow down the focus, say to terrain features. One direction I am willing to explore is model compression techniques, like pruning and quantization (and many others). This not only makes the model smaller, but also faster for inference and very much touches demoscene size-coding spirit. 64kb might still tight though ;P
Sounds cool! Hope you can put it into practice on the big screen soon :)
