Reproduce “Towards Data-Efficient, Real-Time World Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Raluca Stevenson, David Bignell, Jennifer C. White, Tabish Rashid, Emanuele Aiello, Mariya Hendriksen, Junha Hyung, Sadegh Aliakbarian and 4 more
WHAM-RT generates action-conditioned gameplay video by drafting each frame’s image tokens in parallel with a large model, then refining them in a few passes using a lightweight model and compressed gameplay context. In Bleeding Edge, it reaches up to 17 frames per second, reported as 79 times faster than WHAM. A 28-hour model improves long-horizon FVD over full-data WHAM-RT, but remains behind WHAM and covers one character.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.