Motivation
Reviewer btN3: "Including CaR in the camera-following diagnostic would help interpret the closest comparison."
Reviewer btN3: "...do the external-comparison conclusions remain consistent when returns are matched or stratified by realized motion error...?"
Camera following
CaR follows the path as well as we do. The other four models are clearly worse.
Table 1: Camera following by system (five rollout seeds; direction agreement, correlation with the prescribed yaw, implied FOV).
| system | direction | correlation | implied FOV (target 90°) |
|---|---|---|---|
| Memorizon 100 / 200 / 400 s | 0.987 / 0.986 / 0.986 | 0.956 / 0.958 / 0.955 | 90.4 / 90.2 / 90.2 |
| CaR | 0.990 | 0.959 | 60.2 |
| LingBot-World-v2 | 0.976 | 0.879 | 72.5 |
| Matrix-Game 3.0 | 0.971 | 0.866 | 98.1 |
| Infinite-World | 0.961 | 0.888 | 108.6 |
| DreamX-World 1.0 | 0.952 | 0.815 | 133.4 |
| HY-World 1.5 | 0.906 | 0.696 | 123.6 |
CaR's 60° implied FOV points to a narrower camera, not a worse path. First 60 s, five rollout seeds.
Examples
Same first frame and camera path, first 30 s, one rollout seed.
Walk 05
Walk 13
Walk 15
Returns at matched yaw error
Among well-aligned returns, CaR ties Memorizon 100 / 400 s and trails 200 s by about 0.06 Revisit.
Table 2: Return quality (Revisit, mean over five seeds) and Gain over the no-bank control, by system.
| system | Revisit (mean ± seed std) | Gain | pairs |
|---|---|---|---|
| Memorizon 100 s | 0.441 ± 0.018 | 0.305 | 536 |
| Memorizon 200 s | 0.483 ± 0.010 | 0.361 | 457 |
| Memorizon 400 s | 0.435 ± 0.022 | 0.310 | 443 |
| CaR | 0.425 ± 0.016 | 0.289 | 592 |
| HY-World 1.5 | 0.329 ± 0.039 | 0.146 | 92 |
| Matrix-Game 3.0 | 0.302 ± 0.043 | 0.104 | 108 |
| Infinite-World | 0.275 ± 0.026 | 0.164 | 140 |
| LingBot-World-v2 | 0.163 ± 0.032 | 0.006 | 55 |
| DreamX-World 1.0 | 0.139 ± 0.039 | 0.029 | 113 |
Returns with yaw error ≤ 10°, each video calibrated to the prescribed turn amplitude. Yaw comes from phase correlation on the generated video: it sees rotation, not translation.
- Most returns of the other systems fall above 25° error, the confound the reviewer raised.
Conclusion
- CaR follows the path as well as Memorizon (correlation 0.959 vs 0.955–0.958); its 60° implied FOV points to a narrower camera.
- At matched yaw error CaR is within seed noise of Memorizon 100 / 400 s (0.425 vs 0.441 / 0.435) and 0.06 Revisit below 200 s (0.483). HY-World, Matrix-Game, Infinite-World, LingBot and DreamX stay far lower.
- What changes in the paper: camera following moves into Sec. 4.2; the cross-system ranking is called indicative; within-backbone ablations are the primary evidence for memory; no clear margin over CaR is claimed.