A training-free method fixes revisit inconsistency in generative rendering

Wenchao Ma · hf · 2026-07-27

The paper targets revisit inconsistency in autoregressive generative rendering, where a model regenerates different appearances after the camera returns to a previously seen location.
The issue arises because long-horizon generation relies on a bounded KV cache, so earlier context gets evicted.
The proposed training-free method uses correspondences already available from the 3D engine.
Temporal correspondence retrieves pose-matched latent chunks into the KV cache as loop-closure memory, while spatial correspondence uses camera pose and depth reprojection to bias attention toward geometrically matched regions.
On loop-closure trajectories mined from TartanAir and TartanGround, the method improves revisit consistency without hurting overall video quality.

More from Research