Do camera embeddings help when geometry is missing?

Paired comparison

Videos play silently in sync. Proxy previews and generated videos use the same target trajectory. Compare scored views below against ground truth.

Matched scored keyframes

This scene and seed · 12 scored keyframes

All six scenes · mean of two seeds

PSNR / SSIM higher is better; LPIPS / FID / rotation error lower is better. FID is the mean of the two per-seed FIDs, each using 72 keyframes; it is not pooled FID. This small correlated sample makes FID descriptive.

Camera on minus off · scene-paired differences

Average the two seeds within each scene, then bootstrap six scenes (10,000 samples). Intervals are exploratory percentile 95% intervals, without multiple-comparison correction. Negative rotation/LPIPS differences favor camera-on; positive PSNR differences favor camera-on.

Protocol and interpretation

Same release weights, 9 context images, 124 frames at 640 × 480, dpmpp2m with 32 steps, TP4, seeds 42 and 43. No fixed first-view latent. Scores use keyframes 1–12 on the common 512 × 384 canvas; the anchor is excluded. LPIPS uses VGG. Camera-following uses Pi3X poses aligned to the context cameras, not fitted to the generated target path.

For missing geometry, all 3D objects (including layout surfaces) are sorted by ID and randomly permuted with seed 20261110. Nested sets of floor(50%) and floor(75%) of objects are removed and the proxy is rerendered. The original background, contexts, cameras and prompts remain fixed. Percentages refer to object counts, not pixel area. Actual changed-pixel fractions are reported above. Both Qwen and VAE consume the degraded proxy.

Camera-off removes camera encoder/embedding contributions at inference. It still receives proxy motion and the same trajectory-derived text prompt, so this does not remove every source of camera information. Synthetic object removal is not equivalent to naturally failed reconstruction. Three scenes were chosen by largest camera rotation, then three by lowest context coverage among remaining distinct scenes, before generating results.