TinyWorlds

TinyWorlds

Minimal implementation of DeepMind's Genie

Description

World models are hot, but DeepMind's Genie paper leaves you unsure how to code it. TinyWorlds is a minimal autoregressive world model built on the Genie architecture that infers actions between frames from action-less video.

The compact code is ideal for learning how world models work and experimenting on your own data.

Features



Genie architecture:Latent action inference.

No action labels:Learns from video.

Minimal:Easy to read.

Trainable:Your own data.
Tags:llm