Distribution Recovery in Compact Diffusion World Models via Conditioned Frame Interpolation

Abstract

This early proof-of-concept explores addressing distribution drift in diffusion-based world models without requiring massive model scale or constrained environments. We explore a dual-purpose training approach where models learn both autoregressive world generation and frame interpolation capabilities. This is combined with an out-of-distribution detection mechanism that, upon detecting drift or degradation, samples appropriate target frames and conditions the model to interpolate toward them, effectively pulling generation back into the learned distribution. We demonstrate this approach's potential through initial experiments and discuss practical considerations for target frame sampling and interpolation training. This early work presents an alternative path toward enabling longer world exploration with smaller models.