417k-param RNN generates all 6,573 frames of Bad Apple from a single initial state

SEBADA321 · reddit · 2026-09-08

An open-source project trains a tiny recurrent dynamical system (417k params, 1.6MB FP32) to autonomously roll out the full 6,573-frame Bad Apple video from a single 64-dim initial state, with no timestamp input at inference. Key training tricks: learned teacher tables for parallel segment training, rollout-horizon curriculum (K=2 to 512), state perturbation noise, second-difference smoothness regularization, and chunked decoding. It runs at >200 FPS on an RTX 4080 with 17.2MB VRAM and successfully extrapolates far beyond its 512-frame training horizon. Code and weights on GitHub.

Original post →

More from Research

Research channel →