Zero123Collapse

Visual experiments using a pretrained Zero123 model in a recursive generation loop.

Associated code can be found here.

These first experiments are simply an iterative rotation on the azimuth axis of the previous generation.

Zero-123 is heavily biased, and mainly used for geometric optimization rather than outright image generation. Compounded error into collapse was expected, but nowhere near that fast (2-5 frames). Given that the model is designed to transform the input conditioning image, it never actually collapses to a single mode, making for interesting animations.

This also helps to explain why tools using Zero-123 such as Dreamfusion generate such saturated and canonical/simplistic output.