Response to
Reviewer KgbX
Videos for the real-capture test are on the supplementary video page.
Q1 (abstract notation)
Fixed: "bounded by k x K"; the preceding sentence is split for readability.
Q2 (real-world video)
Training on real video needs poses: metric scale is ambiguous in estimated poses, and pose noise changes which frames are retrieved (our retrieval uses pose only), so we keep real-video training as future work. We did add a zero-shot test on real capture: 20 handheld scenes from DL3DV-10K, each given only its first frame and its own camera path (translations rescaled to the training speed, intrinsics fixed at 90 deg horizontal FOV), generated for 10-13 s with the 200 s and 400 s models, which never saw real video. The camera follows the prescribed path (direction agreement 0.93 / 0.90, correlation with the prescribed yaw 0.79 / 0.73 for 400 s / 200 s). Pixel agreement with the real frames is not better than repeating the first frame (PSNR 11.7 vs 12.0, LPIPS 0.58 vs 0.55 over the first 3 s), so we claim path following, not fidelity. The implied horizontal FOV (111-118 deg vs the 90 deg we assume) is unexplained. All 20 clips are on the supplementary video page.
Consistency vs quality
Table 13 attributes the quality loss to the train/test mismatch of the bank (clean at training, self-generated at inference), not to a weaker generator. [PENDING: G1] reports whether bank noise narrows it.