PVWander Rebuttal Supplement

Section 1. SSIM–Perceptual Quality Trade-off (Table 3)

1.1 Frame-level Comparison

Ground truth, a heavily smoothed Online Memory output, and a more detailed but imperfectly aligned Gating output at frame 561.
Fig. 1. Frame 561 comparison.

Gating retains more recognizable scene content and improves LPIPS, yet the smoother Online Memory output receives higher SSIM.

1.2 Controlled Blur Diagnostic

We take the Gating output frame shown in Fig. 1 and apply Gaussian blur (σ = 2 px) directly to the generated image as a post-processing step. Comparing this blurred image with the same ground-truth reference raises SSIM from 0.6010 to 0.6332, while worsening LPIPS from 0.4577 to 0.5781.

Original baseline, original gating, and artificially blurred gating output, with SSIM and LPIPS values.
Fig. 2. Controlled blur diagnostic. Artificially blurred gating output: higher SSIM, less detail, worse LPIPS.

1.3 Full-sequence Comparison

Video 1. Full 641-frame sequence · 30 fps. Left to right: GT / Online Memory / + Timestep-aware Gating. Per-frame SSIM, PSNR and LPIPS are computed on the original images.
Table 1. Full-sequence comparison.
MethodSSIM ↑PSNR ↑LPIPS ↓
Online Memory0.642115.360.5139
+ Timestep-aware Gating0.628716.050.4255

Across all 641 frames, gating improves PSNR and LPIPS, while SSIM drops. Fig. 1 and Fig. 2 illustrate how smoothing can favor SSIM despite losing perceptual detail.

Section 2. Long-horizon Quality over 16 Segments

Segment-wise scores averaged over the same 20 scenes. Each point summarizes one segment: 81 frames for segment 1, then 80 new frames per segment.

2.1 Segment-wise SSIM

SSIM across 16 segments for six models, averaged over the same 20 scenes.
Fig. 3. Segment-wise SSIM over 16 segments. Higher is better.

2.2 Segment-wise PSNR

PSNR across 16 segments for six models, averaged over the same 20 scenes.
Fig. 4. Segment-wise PSNR over 16 segments. Higher is better.

Shared boundary frames are counted once.

Section 4. Qualitative Comparisons