AlayaWorld: Interactive Long-Horizon World Modeling -- Full Technical Report
AlayaWorld claims 24-fps interactive video worlds at 540p and 720p, with four-step inference per chunk.
The system is built on a 15B video diffusion transformer and generates short latent chunks under camera trajectories and switchable text prompts. Its memory design keeps a sink frame, compressed history, geometry-aligned spatial memory, and recent-frame conditioning to hold consistency over time. The team says training on corrupted histories and its own roll-outs reduces long-term drift. On iWorld-Bench, it reports the best long-horizon generation performance. HF Daily Papers' note
The system is built on a 15B video diffusion transformer and generates short latent chunks under camera trajectories and switchable text prompts. Its memory design keeps a sink frame, compressed history, geometry-aligned spatial memory, and recent-frame conditioning to hold consistency over time. The team says training on corrupted histories and its own roll-outs reduces long-term drift. On iWorld-Bench, it reports the best long-horizon generation performance. HF Daily Papers' note
score 6