Closing the Loop: Training-Free Revisit Consistency for Autoregressive Generative Rendering
Wenchao Ma, Changran Liu, Sharon X. Huang, Haomiao Jiang
cs.CV
2026-07-24
Autoregressive video flickers when the camera revisits a place; this training-free method retrieves pose-matched history into the KV cache and biases attention by depth reprojection, nearly doubling keypoint matches without quality loss.
Conditional video generators can turn 3D-engine output (depth maps, untextured geometry) into photorealistic video for games and immersive content. These uses demand long-horizon autoregressive generation: keep synthesizing chunk by chunk while preserving one consistent 3D world.
The problem is that an autoregressive generator with a bounded KV cache evicts old context once the cache fills. When the camera swings back to a place it already visited, the model has lost that memory and has to re-imagine appearance from the conditioning (depth, which still aligns with geometry). The same wall or object ends up with different textures, colors, or structure. The conditioning is fine; the generation is inconsistent.
The authors use a training-free method (no post-training) that reuses correspondences the 3D engine already provides, in the spirit of loop closure from SLAM.
The key design choice is to bias attention logits only, never warping or blending the cached keys, values, or latents, which would smear the cached content. The method rides on Causal Wan-VACE, the depth-conditioned Wan video generator.
Tested on revisit trajectories mined from TartanGround and TartanAir, against four training-free baselines (Self-Forcing, Infinity-RoPE, MemRoPE, Deep Forcing).
| Dataset | Metric | Self-Forcing | Ours |
| TartanGround | high-confidence keypoint matches | 23.26 | 49.08 (nearly doubled) |
| TartanGround | DINO similarity | 0.6087 | 0.6904 |
| TartanAir | keypoint matches | 195.65 | 284.92 |
| TartanAir | DINO similarity | 0.7623 | 0.8407 |
Overall video quality (VBench-Long) is also best, with no quality sacrificed for consistency. In ablation, the attention sink stabilizes global appearance against drift, and pose retrieval is the dominant factor for revisit consistency.
Revisit inconsistency is a concrete, annoying failure in long-horizon autoregressive video, and this gives a no-retraining fix of direct use to teams building real-time game rendering and long video generation. It also demonstrates the more general idea of using whatever correspondences the engine provides to constrain the generator.
It depends on engine-supplied camera poses and metric depth. For real-world video these must be estimated (e.g. with VGGT), and estimation noise bleeds into the correspondences. The authors say the soft Gaussian bias is designed to tolerate such noise, but they do not validate it on real video. It is built only on Wan-VACE, and whether it transfers to other autoregressive video models is unclear.