Same-pose return frame did it come back to what it saw? left = the method's GENERATED frame at the return pose, right = the original OBSERVED GT frame at that same pose — they should match.

Shape-of-Motiongeometric 4D-GS

return-PSNR (full-frame) 18.0 dB · LPIPS 0.560 · degrade 1.0 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 69.6 · photo ↑ 0.0 (EPE 1.68px) · motion 18.87 px

GEN3C3D-cache

return-PSNR (full-frame) 16.2 dB · LPIPS 0.412 · degrade 13.7 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 6.5 · photo ↑ 19.7 (EPE 0.96px) · motion 5.73 px

Lyra-2learned

return-PSNR (full-frame) 12.7 dB · LPIPS 0.535 · degrade 28.1 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

SpMemspatial-mem

return-PSNR (full-frame) 10.6 dB · LPIPS 0.802 · degrade 30.3 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 68.5 · photo ↑ 0.0 (EPE 1.78px) · motion 7.52 px

Ours VACE-14B (revisit-FT)control-only FT

return-PSNR (full-frame) 10.5 dB · LPIPS 0.732 · degrade 30.4 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 59.2 · photo ↑ 0.0 (EPE 2.75px) · motion 9.15 px

MatrixGame-3learned

return-PSNR (full-frame) 10.0 dB · LPIPS 0.634 · degrade 2.0 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Ours VACE-14Bwarp+inpaint

return-PSNR (full-frame) 9.9 dB · LPIPS 0.689 · degrade 30.9 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 56.2 · photo ↑ 0.0 (EPE 1.25px) · motion 6.95 px

Ours VACE-14B (revisit-FT, DiT+VACE-full)DiT+VACE full FT

return-PSNR (full-frame) 8.9 dB · LPIPS 0.764 · degrade 31.9 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 24.9 · photo ↑ 0.0 (EPE 4.80px) · motion 8.59 px

ViewCraftersingle-shot

return-PSNR (full-frame) 12.2 dB · LPIPS 0.667 · degrade 27.7 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 47.7 · photo ↑ 0.0 (EPE 2.80px) · motion 9.62 px

FrameCraftersingle-shot

return-PSNR (full-frame) 9.2 dB · LPIPS 0.745 · degrade 31.6 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

NVS-Solversingle-shot

return-PSNR (full-frame) 8.3 dB · LPIPS 0.890 · degrade 32.5 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Higher return-PSNR = the model came back to the same view it originally observed. degrade = drop from the observed-frame PSNR baseline (how much memory was lost on return). Numbers reused from the moving-back return-fidelity metrics. Dynamic-region PSNR scores the mover pixels only (spatialvid dyn-masks / kubric segmentation, background & >35%-area static planes excluded); it is the HONEST dynamic test — the full-frame number is inflated for warp/copy methods that reproduce the static background perfectly. DL3DV has no mover masks → n/a.

WorldScore consistency ⟳ is reference-free — computed from each method's own generated frames only (no GT poses/depth/frames; the requested camera path is NOT checked). 3D↑ = DROID dense-BA reprojection self-consistency (1−e/1.0719, ×100) — explicit-geometry methods (GEN3C 3D-cache, warp/4D-cache) score highest; a real DROID tracking-loss is recorded as an honest 0. photo↑ = SEA-RAFT forward/backward flow-cycle stability (1−EPE/1.192, ×100); this bound is tuned for smooth single-scene static clips, so under the large inter-frame motion of these revisit clips it saturates near 0 — the raw EPE (px) is shown alongside as the discriminative number. motion = median flow (px), the guard-rail: a frozen/degenerate clip fakes ~100 consistency at ~0 motion, so a high consistency next to near-zero motion is a frozen-cheat, not real memory. Scored on the shown clips for {base VACE-14B, revisit-FT, DiT+VACE-full, 4D-cache} + {GEN3C, Shape-of-Motion, SpMem, ViewCrafter}; other methods show “not scored”.

Same-pose return frame did it come back to what it saw? left = the method's GENERATED frame at the return pose, right = the original OBSERVED GT frame at that same pose — they should match.

Shape-of-Motiongeometric 4D-GS

return-PSNR (full-frame) 16.0 dB · LPIPS 0.686 · degrade 0.4 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 29.2 · photo ↑ 80.8 (EPE 0.23px) · motion 3.47 px

Lyra-2learned

return-PSNR (full-frame) 14.0 dB · LPIPS 0.525 · degrade 27.7 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

GEN3C3D-cache

return-PSNR (full-frame) 14.0 dB · LPIPS 0.546 · degrade 18.3 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 25.3 · photo ↑ 0.0 (EPE 1.25px) · motion 4.55 px

Ours VACE-14Bwarp+inpaint

return-PSNR (full-frame) 13.2 dB · LPIPS 0.593 · degrade 28.6 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 63.1 · photo ↑ 65.0 (EPE 0.42px) · motion 4.95 px

Ours VACE-14B (revisit-FT, DiT+VACE-full)DiT+VACE full FT

return-PSNR (full-frame) 12.7 dB · LPIPS 0.597 · degrade 29.0 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 53.6 · photo ↑ 62.2 (EPE 0.45px) · motion 5.31 px

Ours VACE-14B (revisit-FT)control-only FT

return-PSNR (full-frame) 12.7 dB · LPIPS 0.592 · degrade 29.0 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 58.0 · photo ↑ 66.2 (EPE 0.40px) · motion 4.56 px

SpMemspatial-mem

return-PSNR (full-frame) 10.2 dB · LPIPS 0.799 · degrade 31.5 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 76.7 · photo ↑ 49.3 (EPE 0.60px) · motion 7.10 px

MatrixGame-3learned

return-PSNR (full-frame) 9.9 dB · LPIPS 0.830 · degrade 3.3 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Voyagerworld-model

metrics pending

WorldScore consistency ⟳ ref-free not scored

ViewCraftersingle-shot

return-PSNR (full-frame) 10.3 dB · LPIPS 0.701 · degrade 30.5 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 4.30px) · motion 13.63 px

FrameCraftersingle-shot

return-PSNR (full-frame) 9.2 dB · LPIPS 0.753 · degrade 32.5 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

NVS-Solversingle-shot

return-PSNR (full-frame) 8.3 dB · LPIPS 0.877 · degrade 33.5 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Higher return-PSNR = the model came back to the same view it originally observed. degrade = drop from the observed-frame PSNR baseline (how much memory was lost on return). Numbers reused from the moving-back return-fidelity metrics. Dynamic-region PSNR scores the mover pixels only (spatialvid dyn-masks / kubric segmentation, background & >35%-area static planes excluded); it is the HONEST dynamic test — the full-frame number is inflated for warp/copy methods that reproduce the static background perfectly. DL3DV has no mover masks → n/a.

WorldScore consistency ⟳ is reference-free — computed from each method's own generated frames only (no GT poses/depth/frames; the requested camera path is NOT checked). 3D↑ = DROID dense-BA reprojection self-consistency (1−e/1.0719, ×100) — explicit-geometry methods (GEN3C 3D-cache, warp/4D-cache) score highest; a real DROID tracking-loss is recorded as an honest 0. photo↑ = SEA-RAFT forward/backward flow-cycle stability (1−EPE/1.192, ×100); this bound is tuned for smooth single-scene static clips, so under the large inter-frame motion of these revisit clips it saturates near 0 — the raw EPE (px) is shown alongside as the discriminative number. motion = median flow (px), the guard-rail: a frozen/degenerate clip fakes ~100 consistency at ~0 motion, so a high consistency next to near-zero motion is a frozen-cheat, not real memory. Scored on the shown clips for {base VACE-14B, revisit-FT, DiT+VACE-full, 4D-cache} + {GEN3C, Shape-of-Motion, SpMem, ViewCrafter}; other methods show “not scored”.

Same-pose return frame did it come back to what it saw? left = the method's GENERATED frame at the return pose, right = the original OBSERVED GT frame at that same pose — they should match.

Shape-of-Motiongeometric 4D-GS

return-PSNR (full-frame) 16.9 dB · LPIPS 0.440 · degrade 2.2 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 79.9 · photo ↑ 30.9 (EPE 0.82px) · motion 21.63 px

GEN3C3D-cache

return-PSNR (full-frame) 14.7 dB · LPIPS 0.380 · degrade 16.4 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 3.7 · photo ↑ 30.0 (EPE 0.83px) · motion 5.45 px

Lyra-2learned

return-PSNR (full-frame) 14.0 dB · LPIPS 0.732 · degrade 24.8 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Ours VACE-14B (revisit-FT)control-only FT

return-PSNR (full-frame) 13.4 dB · LPIPS 0.730 · degrade 25.4 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 22.8 · photo ↑ 0.0 (EPE 3.67px) · motion 15.62 px

Ours VACE-14Bwarp+inpaint

return-PSNR (full-frame) 12.6 dB · LPIPS 0.733 · degrade 26.2 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 64.2 · photo ↑ 0.0 (EPE 3.06px) · motion 17.28 px

Ours VACE-14B (revisit-FT, DiT+VACE-full)DiT+VACE full FT

return-PSNR (full-frame) 12.2 dB · LPIPS 0.750 · degrade 26.7 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 17.08px) · motion 15.16 px

MatrixGame-3learned

return-PSNR (full-frame) 10.8 dB · LPIPS 0.731 · degrade 1.2 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

SpMemspatial-mem

return-PSNR (full-frame) 10.5 dB · LPIPS 0.805 · degrade 28.3 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 60.0 · photo ↑ 0.0 (EPE 2.74px) · motion 13.52 px

Voyagerworld-model

metrics pending

WorldScore consistency ⟳ ref-free not scored

ViewCraftersingle-shot

return-PSNR (full-frame) 11.9 dB · LPIPS 0.670 · degrade 26.0 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 47.0 · photo ↑ 40.6 (EPE 0.71px) · motion 9.17 px

FrameCraftersingle-shot

return-PSNR (full-frame) 9.8 dB · LPIPS 0.747 · degrade 29.0 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

NVS-Solversingle-shot

return-PSNR (full-frame) 8.4 dB · LPIPS 0.712 · degrade 30.4 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Higher return-PSNR = the model came back to the same view it originally observed. degrade = drop from the observed-frame PSNR baseline (how much memory was lost on return). Numbers reused from the moving-back return-fidelity metrics. Dynamic-region PSNR scores the mover pixels only (spatialvid dyn-masks / kubric segmentation, background & >35%-area static planes excluded); it is the HONEST dynamic test — the full-frame number is inflated for warp/copy methods that reproduce the static background perfectly. DL3DV has no mover masks → n/a.

WorldScore consistency ⟳ is reference-free — computed from each method's own generated frames only (no GT poses/depth/frames; the requested camera path is NOT checked). 3D↑ = DROID dense-BA reprojection self-consistency (1−e/1.0719, ×100) — explicit-geometry methods (GEN3C 3D-cache, warp/4D-cache) score highest; a real DROID tracking-loss is recorded as an honest 0. photo↑ = SEA-RAFT forward/backward flow-cycle stability (1−EPE/1.192, ×100); this bound is tuned for smooth single-scene static clips, so under the large inter-frame motion of these revisit clips it saturates near 0 — the raw EPE (px) is shown alongside as the discriminative number. motion = median flow (px), the guard-rail: a frozen/degenerate clip fakes ~100 consistency at ~0 motion, so a high consistency next to near-zero motion is a frozen-cheat, not real memory. Scored on the shown clips for {base VACE-14B, revisit-FT, DiT+VACE-full, 4D-cache} + {GEN3C, Shape-of-Motion, SpMem, ViewCrafter}; other methods show “not scored”.

Same-pose return frame did it come back to what it saw? left = the method's GENERATED frame at the return pose, right = the original OBSERVED GT frame at that same pose — they should match.

GEN3C3D-cache

return-PSNR (full-frame) 23.5 dB · LPIPS 0.154 · degrade 11.5 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 28.1 · photo ↑ 47.4 (EPE 0.63px) · motion 4.55 px

Shape-of-Motiongeometric 4D-GS

return-PSNR (full-frame) 19.7 dB · LPIPS 0.381 · degrade 0.6 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 64.4 · photo ↑ 0.0 (EPE 3.62px) · motion 12.19 px

SpMemspatial-mem

return-PSNR (full-frame) 12.8 dB · LPIPS 0.657 · degrade 28.8 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 45.4 · photo ↑ 78.0 (EPE 0.26px) · motion 4.58 px

Ours VACE-14Bwarp+inpaint

return-PSNR (full-frame) 12.5 dB · LPIPS 0.549 · degrade 29.1 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 51.0 · photo ↑ 70.9 (EPE 0.35px) · motion 4.33 px

Lyra-2learned

return-PSNR (full-frame) 11.8 dB · LPIPS 0.644 · degrade 29.8 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Ours VACE-14B (revisit-FT, DiT+VACE-full)DiT+VACE full FT

return-PSNR (full-frame) 11.8 dB · LPIPS 0.632 · degrade 29.9 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 36.7 · photo ↑ 72.5 (EPE 0.33px) · motion 4.12 px

Ours VACE-14B (revisit-FT)control-only FT

return-PSNR (full-frame) 11.4 dB · LPIPS 0.599 · degrade 30.3 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 43.1 · photo ↑ 66.2 (EPE 0.40px) · motion 3.69 px

MatrixGame-3learned

return-PSNR (full-frame) 8.0 dB · LPIPS 0.768 · degrade 3.4 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Voyagerworld-model

metrics pending

WorldScore consistency ⟳ ref-free not scored

NVS-Solversingle-shot

return-PSNR (full-frame) 10.6 dB · LPIPS 0.752 · degrade 31.1 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

FrameCraftersingle-shot

return-PSNR (full-frame) 10.2 dB · LPIPS 0.704 · degrade 31.5 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

ViewCraftersingle-shot

return-PSNR (full-frame) 9.8 dB · LPIPS 0.700 · degrade 30.3 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 52.6 · photo ↑ 46.2 (EPE 0.64px) · motion 5.18 px

Higher return-PSNR = the model came back to the same view it originally observed. degrade = drop from the observed-frame PSNR baseline (how much memory was lost on return). Numbers reused from the moving-back return-fidelity metrics. Dynamic-region PSNR scores the mover pixels only (spatialvid dyn-masks / kubric segmentation, background & >35%-area static planes excluded); it is the HONEST dynamic test — the full-frame number is inflated for warp/copy methods that reproduce the static background perfectly. DL3DV has no mover masks → n/a.

WorldScore consistency ⟳ is reference-free — computed from each method's own generated frames only (no GT poses/depth/frames; the requested camera path is NOT checked). 3D↑ = DROID dense-BA reprojection self-consistency (1−e/1.0719, ×100) — explicit-geometry methods (GEN3C 3D-cache, warp/4D-cache) score highest; a real DROID tracking-loss is recorded as an honest 0. photo↑ = SEA-RAFT forward/backward flow-cycle stability (1−EPE/1.192, ×100); this bound is tuned for smooth single-scene static clips, so under the large inter-frame motion of these revisit clips it saturates near 0 — the raw EPE (px) is shown alongside as the discriminative number. motion = median flow (px), the guard-rail: a frozen/degenerate clip fakes ~100 consistency at ~0 motion, so a high consistency next to near-zero motion is a frozen-cheat, not real memory. Scored on the shown clips for {base VACE-14B, revisit-FT, DiT+VACE-full, 4D-cache} + {GEN3C, Shape-of-Motion, SpMem, ViewCrafter}; other methods show “not scored”.

Same-pose return frame did it come back to what it saw? left = the method's GENERATED frame at the return pose, right = the original OBSERVED GT frame at that same pose — they should match.

Shape-of-Motiongeometric 4D-GS

return-PSNR (full-frame) 20.2 dB · LPIPS 0.505 · degrade 3.1 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 32.4 · photo ↑ 50.3 (EPE 0.59px) · motion 8.28 px

Ours VACE-14B (revisit-FT)control-only FT

return-PSNR (full-frame) 16.5 dB · LPIPS 0.820 · degrade 26.3 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 31.8 · photo ↑ 24.5 (EPE 0.90px) · motion 7.42 px

GEN3C3D-cache

return-PSNR (full-frame) 16.4 dB · LPIPS 0.531 · degrade 20.2 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 26.2 · photo ↑ 7.2 (EPE 1.11px) · motion 9.74 px

Ours VACE-14B (revisit-FT, DiT+VACE-full)DiT+VACE full FT

return-PSNR (full-frame) 15.8 dB · LPIPS 0.668 · degrade 27.0 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 45.8 (EPE 0.65px) · motion 8.54 px

Lyra-2learned

return-PSNR (full-frame) 15.4 dB · LPIPS 0.643 · degrade 27.5 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Ours VACE-14Bwarp+inpaint

return-PSNR (full-frame) 14.6 dB · LPIPS 0.648 · degrade 28.2 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 25.0 · photo ↑ 56.1 (EPE 0.52px) · motion 8.00 px

MatrixGame-3learned

return-PSNR (full-frame) 13.8 dB · LPIPS 0.698 · degrade 1.7 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

SpMemspatial-mem

return-PSNR (full-frame) 11.2 dB · LPIPS 0.942 · degrade 31.7 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 19.7 · photo ↑ 0.0 (EPE 1.77px) · motion 8.02 px

Voyagerworld-model

metrics pending

WorldScore consistency ⟳ ref-free not scored

ViewCraftersingle-shot

return-PSNR (full-frame) 15.2 dB · LPIPS 0.738 · degrade 26.5 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 34.2 · photo ↑ 0.0 (EPE 1.61px) · motion 9.54 px

FrameCraftersingle-shot

return-PSNR (full-frame) 14.4 dB · LPIPS 0.621 · degrade 28.4 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

NVS-Solversingle-shot

return-PSNR (full-frame) 11.4 dB · LPIPS 0.802 · degrade 31.5 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Higher return-PSNR = the model came back to the same view it originally observed. degrade = drop from the observed-frame PSNR baseline (how much memory was lost on return). Numbers reused from the moving-back return-fidelity metrics. Dynamic-region PSNR scores the mover pixels only (spatialvid dyn-masks / kubric segmentation, background & >35%-area static planes excluded); it is the HONEST dynamic test — the full-frame number is inflated for warp/copy methods that reproduce the static background perfectly. DL3DV has no mover masks → n/a.

WorldScore consistency ⟳ is reference-free — computed from each method's own generated frames only (no GT poses/depth/frames; the requested camera path is NOT checked). 3D↑ = DROID dense-BA reprojection self-consistency (1−e/1.0719, ×100) — explicit-geometry methods (GEN3C 3D-cache, warp/4D-cache) score highest; a real DROID tracking-loss is recorded as an honest 0. photo↑ = SEA-RAFT forward/backward flow-cycle stability (1−EPE/1.192, ×100); this bound is tuned for smooth single-scene static clips, so under the large inter-frame motion of these revisit clips it saturates near 0 — the raw EPE (px) is shown alongside as the discriminative number. motion = median flow (px), the guard-rail: a frozen/degenerate clip fakes ~100 consistency at ~0 motion, so a high consistency next to near-zero motion is a frozen-cheat, not real memory. Scored on the shown clips for {base VACE-14B, revisit-FT, DiT+VACE-full, 4D-cache} + {GEN3C, Shape-of-Motion, SpMem, ViewCrafter}; other methods show “not scored”.

Same-pose return frame did it come back to what it saw? left = the method's GENERATED frame at the return pose, right = the original OBSERVED GT frame at that same pose — they should match.

Shape-of-Motiongeometric 4D-GS

return-PSNR (full-frame) 28.0 dB · LPIPS 0.270 · degrade -0.4 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 17.1 · photo ↑ 64.5 (EPE 0.42px) · motion 23.50 px

Lyra-2learned

return-PSNR (full-frame) 16.5 dB · LPIPS 0.643 · degrade 27.1 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

MatrixGame-3learned

return-PSNR (full-frame) 16.5 dB · LPIPS 0.645 · degrade -0.3 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

GEN3C3D-cache

return-PSNR (full-frame) 16.3 dB · LPIPS 0.689 · degrade 21.4 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 38.9 · photo ↑ 0.0 (EPE 1.40px) · motion 15.48 px

Ours VACE-14B (revisit-FT)control-only FT

return-PSNR (full-frame) 16.2 dB · LPIPS 0.684 · degrade 27.4 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 5.06px) · motion 27.04 px

Ours VACE-14B (revisit-FT, DiT+VACE-full)DiT+VACE full FT

return-PSNR (full-frame) 15.4 dB · LPIPS 0.735 · degrade 28.2 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 13.6 · photo ↑ 0.0 (EPE 17.38px) · motion 32.94 px

Ours VACE-14Bwarp+inpaint

return-PSNR (full-frame) 15.0 dB · LPIPS 0.841 · degrade 28.7 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 28.9 · photo ↑ 0.0 (EPE 2.26px) · motion 25.31 px

SpMemspatial-mem

return-PSNR (full-frame) 13.4 dB · LPIPS 0.741 · degrade 30.2 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 10.5 · photo ↑ 0.0 (EPE 1.62px) · motion 18.45 px

Voyagerworld-model

metrics pending

WorldScore consistency ⟳ ref-free not scored

ViewCraftersingle-shot

return-PSNR (full-frame) 15.8 dB · LPIPS 0.562 · degrade 27.1 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 41.3 · photo ↑ 0.0 (EPE 1.44px) · motion 27.86 px

NVS-Solversingle-shot

return-PSNR (full-frame) 11.4 dB · LPIPS 0.784 · degrade 32.2 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

FrameCraftersingle-shot

return-PSNR (full-frame) 11.3 dB · LPIPS 0.783 · degrade 32.3 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Higher return-PSNR = the model came back to the same view it originally observed. degrade = drop from the observed-frame PSNR baseline (how much memory was lost on return). Numbers reused from the moving-back return-fidelity metrics. Dynamic-region PSNR scores the mover pixels only (spatialvid dyn-masks / kubric segmentation, background & >35%-area static planes excluded); it is the HONEST dynamic test — the full-frame number is inflated for warp/copy methods that reproduce the static background perfectly. DL3DV has no mover masks → n/a.

WorldScore consistency ⟳ is reference-free — computed from each method's own generated frames only (no GT poses/depth/frames; the requested camera path is NOT checked). 3D↑ = DROID dense-BA reprojection self-consistency (1−e/1.0719, ×100) — explicit-geometry methods (GEN3C 3D-cache, warp/4D-cache) score highest; a real DROID tracking-loss is recorded as an honest 0. photo↑ = SEA-RAFT forward/backward flow-cycle stability (1−EPE/1.192, ×100); this bound is tuned for smooth single-scene static clips, so under the large inter-frame motion of these revisit clips it saturates near 0 — the raw EPE (px) is shown alongside as the discriminative number. motion = median flow (px), the guard-rail: a frozen/degenerate clip fakes ~100 consistency at ~0 motion, so a high consistency next to near-zero motion is a frozen-cheat, not real memory. Scored on the shown clips for {base VACE-14B, revisit-FT, DiT+VACE-full, 4D-cache} + {GEN3C, Shape-of-Motion, SpMem, ViewCrafter}; other methods show “not scored”.

Same-pose return frame did it come back to what it saw? left = the method's GENERATED frame at the return pose, right = the original OBSERVED GT frame at that same pose — they should match.

GEN3C3D-cache

return-PSNR (full-frame) 18.1 dB · LPIPS 0.357 · degrade 18.9 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 1.9 · photo ↑ 24.2 (EPE 0.90px) · motion 5.55 px

Shape-of-Motiongeometric 4D-GS

return-PSNR (full-frame) 15.0 dB · LPIPS 0.595 · degrade 1.1 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 8.6 · photo ↑ 0.0 (EPE 4.15px) · motion 12.19 px

Ours VACE-14Bwarp+inpaint

return-PSNR (full-frame) 14.4 dB · LPIPS 0.640 · degrade 29.1 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 26.1 · photo ↑ 35.2 (EPE 0.77px) · motion 6.61 px

Ours VACE-14B (revisit-FT)control-only FT

return-PSNR (full-frame) 13.4 dB · LPIPS 0.623 · degrade 30.1 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 13.5 · photo ↑ 41.9 (EPE 0.69px) · motion 6.03 px

Ours VACE-14B (revisit-FT, DiT+VACE-full)DiT+VACE full FT

return-PSNR (full-frame) 12.1 dB · LPIPS 0.643 · degrade 31.4 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 1.30px) · motion 6.49 px

Lyra-2learned

return-PSNR (full-frame) 11.9 dB · LPIPS 0.700 · degrade 31.6 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

MatrixGame-3learned

return-PSNR (full-frame) 11.2 dB · LPIPS 0.802 · degrade -0.7 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

SpMemspatial-mem

return-PSNR (full-frame) 10.1 dB · LPIPS 0.789 · degrade 33.4 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 9.7 · photo ↑ 12.5 (EPE 1.04px) · motion 4.97 px

Voyagerworld-model

metrics pending

WorldScore consistency ⟳ ref-free not scored

ViewCraftersingle-shot

return-PSNR (full-frame) 11.4 dB · LPIPS 0.569 · degrade 31.0 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 16.8 · photo ↑ 30.7 (EPE 0.83px) · motion 6.81 px

FrameCraftersingle-shot

return-PSNR (full-frame) 10.0 dB · LPIPS 0.750 · degrade 33.5 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

NVS-Solversingle-shot

return-PSNR (full-frame) 9.3 dB · LPIPS 0.823 · degrade 34.2 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Higher return-PSNR = the model came back to the same view it originally observed. degrade = drop from the observed-frame PSNR baseline (how much memory was lost on return). Numbers reused from the moving-back return-fidelity metrics. Dynamic-region PSNR scores the mover pixels only (spatialvid dyn-masks / kubric segmentation, background & >35%-area static planes excluded); it is the HONEST dynamic test — the full-frame number is inflated for warp/copy methods that reproduce the static background perfectly. DL3DV has no mover masks → n/a.

WorldScore consistency ⟳ is reference-free — computed from each method's own generated frames only (no GT poses/depth/frames; the requested camera path is NOT checked). 3D↑ = DROID dense-BA reprojection self-consistency (1−e/1.0719, ×100) — explicit-geometry methods (GEN3C 3D-cache, warp/4D-cache) score highest; a real DROID tracking-loss is recorded as an honest 0. photo↑ = SEA-RAFT forward/backward flow-cycle stability (1−EPE/1.192, ×100); this bound is tuned for smooth single-scene static clips, so under the large inter-frame motion of these revisit clips it saturates near 0 — the raw EPE (px) is shown alongside as the discriminative number. motion = median flow (px), the guard-rail: a frozen/degenerate clip fakes ~100 consistency at ~0 motion, so a high consistency next to near-zero motion is a frozen-cheat, not real memory. Scored on the shown clips for {base VACE-14B, revisit-FT, DiT+VACE-full, 4D-cache} + {GEN3C, Shape-of-Motion, SpMem, ViewCrafter}; other methods show “not scored”.

Same-pose return frame did it come back to what it saw? left = the method's GENERATED frame at the return pose, right = the original OBSERVED GT frame at that same pose — they should match.

Shape-of-Motiongeometric 4D-GS

return-PSNR (full-frame) 18.2 dB · LPIPS 0.599 · degrade 0.0 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 51.3 · photo ↑ 0.0 (EPE 1.70px) · motion 30.61 px

Ours VACE-14B (revisit-FT, DiT+VACE-full)DiT+VACE full FT

return-PSNR (full-frame) 15.9 dB · LPIPS 0.866 · degrade 26.2 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 52.5 · photo ↑ 0.0 (EPE 1.43px) · motion 26.98 px

Ours VACE-14Bwarp+inpaint

return-PSNR (full-frame) 15.7 dB · LPIPS 0.723 · degrade 26.3 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 60.3 · photo ↑ 0.0 (EPE 2.25px) · motion 31.42 px

Ours VACE-14B (revisit-FT)control-only FT

return-PSNR (full-frame) 14.5 dB · LPIPS 0.755 · degrade 27.6 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 47.6 · photo ↑ 0.0 (EPE 2.56px) · motion 30.25 px

GEN3C3D-cache

return-PSNR (full-frame) 14.4 dB · LPIPS 0.523 · degrade 20.6 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 28.6 · photo ↑ 6.0 (EPE 1.12px) · motion 10.45 px

SpMemspatial-mem

return-PSNR (full-frame) 14.0 dB · LPIPS 0.720 · degrade 28.1 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 46.4 · photo ↑ 0.0 (EPE 1.31px) · motion 26.88 px

Lyra-2learned

return-PSNR (full-frame) 13.8 dB · LPIPS 0.599 · degrade 28.2 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

MatrixGame-3learned

return-PSNR (full-frame) 9.2 dB · LPIPS 0.856 · degrade 5.3 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Voyagerworld-model

metrics pending

WorldScore consistency ⟳ ref-free not scored

ViewCraftersingle-shot

return-PSNR (full-frame) 13.1 dB · LPIPS 0.581 · degrade 27.8 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 43.3 · photo ↑ 0.0 (EPE 4.33px) · motion 31.38 px

NVS-Solversingle-shot

return-PSNR (full-frame) 12.5 dB · LPIPS 0.830 · degrade 29.6 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

FrameCraftersingle-shot

return-PSNR (full-frame) 11.4 dB · LPIPS 0.712 · degrade 30.6 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Higher return-PSNR = the model came back to the same view it originally observed. degrade = drop from the observed-frame PSNR baseline (how much memory was lost on return). Numbers reused from the moving-back return-fidelity metrics. Dynamic-region PSNR scores the mover pixels only (spatialvid dyn-masks / kubric segmentation, background & >35%-area static planes excluded); it is the HONEST dynamic test — the full-frame number is inflated for warp/copy methods that reproduce the static background perfectly. DL3DV has no mover masks → n/a.

WorldScore consistency ⟳ is reference-free — computed from each method's own generated frames only (no GT poses/depth/frames; the requested camera path is NOT checked). 3D↑ = DROID dense-BA reprojection self-consistency (1−e/1.0719, ×100) — explicit-geometry methods (GEN3C 3D-cache, warp/4D-cache) score highest; a real DROID tracking-loss is recorded as an honest 0. photo↑ = SEA-RAFT forward/backward flow-cycle stability (1−EPE/1.192, ×100); this bound is tuned for smooth single-scene static clips, so under the large inter-frame motion of these revisit clips it saturates near 0 — the raw EPE (px) is shown alongside as the discriminative number. motion = median flow (px), the guard-rail: a frozen/degenerate clip fakes ~100 consistency at ~0 motion, so a high consistency next to near-zero motion is a frozen-cheat, not real memory. Scored on the shown clips for {base VACE-14B, revisit-FT, DiT+VACE-full, 4D-cache} + {GEN3C, Shape-of-Motion, SpMem, ViewCrafter}; other methods show “not scored”.

Same-pose return frame did it come back to what it saw? left = the method's GENERATED frame at the return pose, right = the original OBSERVED GT frame at that same pose — they should match.

GEN3C3D-cache

return-PSNR (full-frame) 27.7 dB · LPIPS 0.140 · degrade 13.0 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 2.97px) · motion 12.57 px

Shape-of-Motiongeometric 4D-GS

return-PSNR (full-frame) 25.1 dB · LPIPS 0.190 · degrade 0.0 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 33.1 · photo ↑ 1.2 (EPE 1.18px) · motion 23.59 px

Lyra-2learned

return-PSNR (full-frame) 19.5 dB · LPIPS 0.333 · degrade 24.6 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Ours VACE-14B (revisit-FT)control-only FT

return-PSNR (full-frame) 18.4 dB · LPIPS 0.663 · degrade 25.7 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 26.6 · photo ↑ 0.0 (EPE 4.31px) · motion 26.08 px

Ours VACE-14B (revisit-FT, DiT+VACE-full)DiT+VACE full FT

return-PSNR (full-frame) 17.5 dB · LPIPS 0.657 · degrade 26.6 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 44.7 · photo ↑ 0.0 (EPE 3.66px) · motion 24.96 px

SpMemspatial-mem

return-PSNR (full-frame) 17.4 dB · LPIPS 0.420 · degrade 26.7 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 6.4 · photo ↑ 0.0 (EPE 2.78px) · motion 22.10 px

Ours VACE-14Bwarp+inpaint

return-PSNR (full-frame) 17.2 dB · LPIPS 0.659 · degrade 26.9 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 37.3 · photo ↑ 0.0 (EPE 3.04px) · motion 24.74 px

MatrixGame-3learned

return-PSNR (full-frame) 16.8 dB · LPIPS 0.615 · degrade 0.3 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

ViewCraftersingle-shot

return-PSNR (full-frame) 19.9 dB · LPIPS 0.537 · degrade 23.3 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 26.8 · photo ↑ 0.0 (EPE 3.23px) · motion 26.91 px

NVS-Solversingle-shot

return-PSNR (full-frame) 14.3 dB · LPIPS 0.488 · degrade 29.7 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

FrameCraftersingle-shot

return-PSNR (full-frame) 13.6 dB · LPIPS 0.763 · degrade 30.5 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Higher return-PSNR = the model came back to the same view it originally observed. degrade = drop from the observed-frame PSNR baseline (how much memory was lost on return). Numbers reused from the moving-back return-fidelity metrics. Dynamic-region PSNR scores the mover pixels only (spatialvid dyn-masks / kubric segmentation, background & >35%-area static planes excluded); it is the HONEST dynamic test — the full-frame number is inflated for warp/copy methods that reproduce the static background perfectly. DL3DV has no mover masks → n/a.

WorldScore consistency ⟳ is reference-free — computed from each method's own generated frames only (no GT poses/depth/frames; the requested camera path is NOT checked). 3D↑ = DROID dense-BA reprojection self-consistency (1−e/1.0719, ×100) — explicit-geometry methods (GEN3C 3D-cache, warp/4D-cache) score highest; a real DROID tracking-loss is recorded as an honest 0. photo↑ = SEA-RAFT forward/backward flow-cycle stability (1−EPE/1.192, ×100); this bound is tuned for smooth single-scene static clips, so under the large inter-frame motion of these revisit clips it saturates near 0 — the raw EPE (px) is shown alongside as the discriminative number. motion = median flow (px), the guard-rail: a frozen/degenerate clip fakes ~100 consistency at ~0 motion, so a high consistency next to near-zero motion is a frozen-cheat, not real memory. Scored on the shown clips for {base VACE-14B, revisit-FT, DiT+VACE-full, 4D-cache} + {GEN3C, Shape-of-Motion, SpMem, ViewCrafter}; other methods show “not scored”.

Same-pose return frame did it come back to what it saw? left = the method's GENERATED frame at the return pose, right = the original OBSERVED GT frame at that same pose — they should match.

Shape-of-Motiongeometric 4D-GS

return-PSNR (full-frame) 20.2 dB · LPIPS 0.457 · degrade 0.5 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 20.3 · photo ↑ 0.0 (EPE 8.86px) · motion 34.73 px

Ours VACE-14B (revisit-FT, DiT+VACE-full)DiT+VACE full FT

return-PSNR (full-frame) 13.1 dB · LPIPS 0.846 · degrade 28.6 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 22.8 · photo ↑ 0.0 (EPE 6.21px) · motion 23.42 px

GEN3C3D-cache

return-PSNR (full-frame) 13.0 dB · LPIPS 0.575 · degrade 21.9 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 21.1 · photo ↑ 0.0 (EPE 2.81px) · motion 17.66 px

Lyra-2learned

return-PSNR (full-frame) 12.8 dB · LPIPS 0.905 · degrade 28.9 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Ours VACE-14Bwarp+inpaint

return-PSNR (full-frame) 12.1 dB · LPIPS 0.745 · degrade 29.5 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 55.1 · photo ↑ 0.0 (EPE 1.43px) · motion 20.46 px

Ours VACE-14B (revisit-FT)control-only FT

return-PSNR (full-frame) 10.7 dB · LPIPS 0.799 · degrade 31.0 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 35.1 · photo ↑ 0.0 (EPE 2.23px) · motion 23.74 px

MatrixGame-3learned

return-PSNR (full-frame) 10.0 dB · LPIPS 0.775 · degrade 1.4 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

SpMemspatial-mem

return-PSNR (full-frame) 9.3 dB · LPIPS 0.786 · degrade 32.4 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 21.1 · photo ↑ 0.0 (EPE 2.76px) · motion 26.16 px

ViewCraftersingle-shot

return-PSNR (full-frame) 11.8 dB · LPIPS 0.678 · degrade 29.0 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 27.2 · photo ↑ 0.0 (EPE 5.90px) · motion 25.62 px

FrameCraftersingle-shot

return-PSNR (full-frame) 9.9 dB · LPIPS 0.745 · degrade 31.8 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

NVS-Solversingle-shot

return-PSNR (full-frame) 9.3 dB · LPIPS 0.953 · degrade 32.4 dB

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Higher return-PSNR = the model came back to the same view it originally observed. degrade = drop from the observed-frame PSNR baseline (how much memory was lost on return). Numbers reused from the moving-back return-fidelity metrics. Dynamic-region PSNR scores the mover pixels only (spatialvid dyn-masks / kubric segmentation, background & >35%-area static planes excluded); it is the HONEST dynamic test — the full-frame number is inflated for warp/copy methods that reproduce the static background perfectly. DL3DV has no mover masks → n/a.

WorldScore consistency ⟳ is reference-free — computed from each method's own generated frames only (no GT poses/depth/frames; the requested camera path is NOT checked). 3D↑ = DROID dense-BA reprojection self-consistency (1−e/1.0719, ×100) — explicit-geometry methods (GEN3C 3D-cache, warp/4D-cache) score highest; a real DROID tracking-loss is recorded as an honest 0. photo↑ = SEA-RAFT forward/backward flow-cycle stability (1−EPE/1.192, ×100); this bound is tuned for smooth single-scene static clips, so under the large inter-frame motion of these revisit clips it saturates near 0 — the raw EPE (px) is shown alongside as the discriminative number. motion = median flow (px), the guard-rail: a frozen/degenerate clip fakes ~100 consistency at ~0 motion, so a high consistency next to near-zero motion is a frozen-cheat, not real memory. Scored on the shown clips for {base VACE-14B, revisit-FT, DiT+VACE-full, 4D-cache} + {GEN3C, Shape-of-Motion, SpMem, ViewCrafter}; other methods show “not scored”.

Same-pose return frames revisit fidelity grouped by pose: each box leads with the OBSERVED GT for that pose (the reference target), then that method's Visit 1 / 2 / 3 of the SAME pose. self-PSNR = cross-visit consistency, GT-PSNR = match to the real observed frame.

Shape-of-Motiongeometric 4D-GS

self-PSNR 62.8 dB · GT-PSNR (full-frame) 13.8 dB · GT-LPIPS 0.820

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 60.4 · photo ↑ 0.0 (EPE 3.06px) · motion 28.81 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Lyra-2learned

self-PSNR 16.7 dB · GT-PSNR (full-frame) 12.4 dB · GT-LPIPS 0.597

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

SpMemspatial-mem

self-PSNR 15.0 dB · GT-PSNR (full-frame) 10.1 dB · GT-LPIPS 0.791

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 20.9 · photo ↑ 0.0 (EPE 5.33px) · motion 11.34 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

GEN3C3D-cache

self-PSNR 13.8 dB · GT-PSNR (full-frame) 14.4 dB · GT-LPIPS 0.524

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 2.31px) · motion 7.83 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Ours VACE-14B (revisit-FT)control-only FT

self-PSNR 13.4 dB · GT-PSNR (full-frame) 13.5 dB · GT-LPIPS 0.630

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 9.43px) · motion 17.34 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Ours VACE-14Bwarp+inpaint

self-PSNR 12.9 dB · GT-PSNR (full-frame) 14.0 dB · GT-LPIPS 0.553

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 14.0 · photo ↑ 0.0 (EPE 5.62px) · motion 15.32 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Ours VACE-14B (revisit-FT, DiT+VACE-full)DiT+VACE full FT

self-PSNR 11.9 dB · GT-PSNR (full-frame) 13.1 dB · GT-LPIPS 0.590

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 5.68px) · motion 16.72 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

MatrixGame-3learned

self-PSNR 10.1 dB · GT-PSNR (full-frame) 9.1 dB · GT-LPIPS 0.768

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

FrameCraftersingle-shot

self-PSNR 16.4 dB · GT-PSNR (full-frame) 15.1 dB · GT-LPIPS 0.461

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

NVS-Solversingle-shot

self-PSNR 12.7 dB · GT-PSNR (full-frame) 11.9 dB · GT-LPIPS 0.770

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

ViewCraftersingle-shot

self-PSNR 11.9 dB · GT-PSNR (full-frame) 11.2 dB · GT-LPIPS 0.675

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 6.47px) · motion 15.51 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Geometric-memory methods (Shape-of-Motion) reproduce the revisited pose near-perfectly (self-PSNR often >50 dB), while learned / single-shot methods drift between visits (self-PSNR ~12–19 dB). Numbers are read directly from the revisit metrics (not recomputed). Dynamic-region PSNR scores the mover pixels only (spatialvid dyn-masks / kubric segmentation, background & >35%-area static planes excluded) — the full-frame GT-PSNR is inflated for warp/copy methods by static-background copying. DL3DV has no masks → n/a.

WorldScore consistency ⟳ is reference-free — computed from each method's own generated frames only (no GT poses/depth/frames; the requested camera path is NOT checked). 3D↑ = DROID dense-BA reprojection self-consistency (1−e/1.0719, ×100) — explicit-geometry methods (GEN3C 3D-cache, warp/4D-cache) score highest; a real DROID tracking-loss is recorded as an honest 0. photo↑ = SEA-RAFT forward/backward flow-cycle stability (1−EPE/1.192, ×100); this bound is tuned for smooth single-scene static clips, so under the large inter-frame motion of these revisit clips it saturates near 0 — the raw EPE (px) is shown alongside as the discriminative number. motion = median flow (px), the guard-rail: a frozen/degenerate clip fakes ~100 consistency at ~0 motion, so a high consistency next to near-zero motion is a frozen-cheat, not real memory. Scored on the shown clips for {base VACE-14B, revisit-FT, DiT+VACE-full, 4D-cache} + {GEN3C, Shape-of-Motion, SpMem, ViewCrafter}; other methods show “not scored”.

Same-pose return frames revisit fidelity grouped by pose: each box leads with the OBSERVED GT for that pose (the reference target), then that method's Visit 1 / 2 / 3 of the SAME pose. self-PSNR = cross-visit consistency, GT-PSNR = match to the real observed frame.

Shape-of-Motiongeometric 4D-GS

self-PSNR 53.5 dB · GT-PSNR (full-frame) 17.5 dB · GT-LPIPS 0.674

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 20.62px) · motion 41.51 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

SpMemspatial-mem

self-PSNR 14.7 dB · GT-PSNR (full-frame) 11.3 dB · GT-LPIPS 0.780

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 6.63px) · motion 24.47 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

GEN3C3D-cache

self-PSNR 13.8 dB · GT-PSNR (full-frame) 14.9 dB · GT-LPIPS 0.507

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 2.29px) · motion 7.13 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Ours VACE-14B (revisit-FT)control-only FT

self-PSNR 12.5 dB · GT-PSNR (full-frame) 12.6 dB · GT-LPIPS 0.631

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 10.34px) · motion 31.99 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

MatrixGame-3learned

self-PSNR 12.1 dB · GT-PSNR (full-frame) 10.7 dB · GT-LPIPS 0.777

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Ours VACE-14Bwarp+inpaint

self-PSNR 11.7 dB · GT-PSNR (full-frame) 12.6 dB · GT-LPIPS 0.621

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 15.92px) · motion 35.12 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Lyra-2learned

self-PSNR 11.5 dB · GT-PSNR (full-frame) 12.1 dB · GT-LPIPS 0.656

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Ours VACE-14B (revisit-FT, DiT+VACE-full)DiT+VACE full FT

self-PSNR 10.6 dB · GT-PSNR (full-frame) 12.1 dB · GT-LPIPS 0.635

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 13.50px) · motion 29.00 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Voyagerworld-model

metrics pending

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

FrameCraftersingle-shot

self-PSNR 18.3 dB · GT-PSNR (full-frame) 16.9 dB · GT-LPIPS 0.364

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

ViewCraftersingle-shot

self-PSNR 12.8 dB · GT-PSNR (full-frame) 10.5 dB · GT-LPIPS 0.656

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 20.12px) · motion 32.74 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

NVS-Solversingle-shot

self-PSNR 9.7 dB · GT-PSNR (full-frame) 10.0 dB · GT-LPIPS 0.773

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Geometric-memory methods (Shape-of-Motion) reproduce the revisited pose near-perfectly (self-PSNR often >50 dB), while learned / single-shot methods drift between visits (self-PSNR ~12–19 dB). Numbers are read directly from the revisit metrics (not recomputed). Dynamic-region PSNR scores the mover pixels only (spatialvid dyn-masks / kubric segmentation, background & >35%-area static planes excluded) — the full-frame GT-PSNR is inflated for warp/copy methods by static-background copying. DL3DV has no masks → n/a.

WorldScore consistency ⟳ is reference-free — computed from each method's own generated frames only (no GT poses/depth/frames; the requested camera path is NOT checked). 3D↑ = DROID dense-BA reprojection self-consistency (1−e/1.0719, ×100) — explicit-geometry methods (GEN3C 3D-cache, warp/4D-cache) score highest; a real DROID tracking-loss is recorded as an honest 0. photo↑ = SEA-RAFT forward/backward flow-cycle stability (1−EPE/1.192, ×100); this bound is tuned for smooth single-scene static clips, so under the large inter-frame motion of these revisit clips it saturates near 0 — the raw EPE (px) is shown alongside as the discriminative number. motion = median flow (px), the guard-rail: a frozen/degenerate clip fakes ~100 consistency at ~0 motion, so a high consistency next to near-zero motion is a frozen-cheat, not real memory. Scored on the shown clips for {base VACE-14B, revisit-FT, DiT+VACE-full, 4D-cache} + {GEN3C, Shape-of-Motion, SpMem, ViewCrafter}; other methods show “not scored”.

Same-pose return frames revisit fidelity grouped by pose: each box leads with the OBSERVED GT for that pose (the reference target), then that method's Visit 1 / 2 / 3 of the SAME pose. self-PSNR = cross-visit consistency, GT-PSNR = match to the real observed frame.

Shape-of-Motiongeometric 4D-GS

self-PSNR 51.0 dB · GT-PSNR (full-frame) 16.9 dB · GT-LPIPS 0.453

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 57.1 · photo ↑ 0.0 (EPE 5.43px) · motion 26.91 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

MatrixGame-3learned

self-PSNR 18.9 dB · GT-PSNR (full-frame) 10.5 dB · GT-LPIPS 0.923

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Ours VACE-14Bwarp+inpaint

self-PSNR 14.6 dB · GT-PSNR (full-frame) 14.5 dB · GT-LPIPS 0.527

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 14.3 · photo ↑ 0.0 (EPE 8.57px) · motion 19.92 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

GEN3C3D-cache

self-PSNR 14.3 dB · GT-PSNR (full-frame) 14.7 dB · GT-LPIPS 0.449

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 19.0 (EPE 0.97px) · motion 5.33 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Ours VACE-14B (revisit-FT)control-only FT

self-PSNR 13.8 dB · GT-PSNR (full-frame) 14.5 dB · GT-LPIPS 0.518

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 7.41px) · motion 20.58 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Lyra-2learned

self-PSNR 12.6 dB · GT-PSNR (full-frame) 12.2 dB · GT-LPIPS 0.628

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

SpMemspatial-mem

self-PSNR 12.2 dB · GT-PSNR (full-frame) 10.8 dB · GT-LPIPS 0.776

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 5.26px) · motion 12.63 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Ours VACE-14B (revisit-FT, DiT+VACE-full)DiT+VACE full FT

self-PSNR 11.6 dB · GT-PSNR (full-frame) 13.5 dB · GT-LPIPS 0.510

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 22.5 · photo ↑ 0.0 (EPE 6.31px) · motion 19.33 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Voyagerworld-model

metrics pending

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

FrameCraftersingle-shot

self-PSNR 16.1 dB · GT-PSNR (full-frame) 15.5 dB · GT-LPIPS 0.337

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

NVS-Solversingle-shot

self-PSNR 12.2 dB · GT-PSNR (full-frame) 10.9 dB · GT-LPIPS 0.678

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

ViewCraftersingle-shot

self-PSNR 11.1 dB · GT-PSNR (full-frame) 11.0 dB · GT-LPIPS 0.661

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 4.06px) · motion 13.89 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Geometric-memory methods (Shape-of-Motion) reproduce the revisited pose near-perfectly (self-PSNR often >50 dB), while learned / single-shot methods drift between visits (self-PSNR ~12–19 dB). Numbers are read directly from the revisit metrics (not recomputed). Dynamic-region PSNR scores the mover pixels only (spatialvid dyn-masks / kubric segmentation, background & >35%-area static planes excluded) — the full-frame GT-PSNR is inflated for warp/copy methods by static-background copying. DL3DV has no masks → n/a.

WorldScore consistency ⟳ is reference-free — computed from each method's own generated frames only (no GT poses/depth/frames; the requested camera path is NOT checked). 3D↑ = DROID dense-BA reprojection self-consistency (1−e/1.0719, ×100) — explicit-geometry methods (GEN3C 3D-cache, warp/4D-cache) score highest; a real DROID tracking-loss is recorded as an honest 0. photo↑ = SEA-RAFT forward/backward flow-cycle stability (1−EPE/1.192, ×100); this bound is tuned for smooth single-scene static clips, so under the large inter-frame motion of these revisit clips it saturates near 0 — the raw EPE (px) is shown alongside as the discriminative number. motion = median flow (px), the guard-rail: a frozen/degenerate clip fakes ~100 consistency at ~0 motion, so a high consistency next to near-zero motion is a frozen-cheat, not real memory. Scored on the shown clips for {base VACE-14B, revisit-FT, DiT+VACE-full, 4D-cache} + {GEN3C, Shape-of-Motion, SpMem, ViewCrafter}; other methods show “not scored”.

Same-pose return frames revisit fidelity grouped by pose: each box leads with the OBSERVED GT for that pose (the reference target), then that method's Visit 1 / 2 / 3 of the SAME pose. self-PSNR = cross-visit consistency, GT-PSNR = match to the real observed frame.

Shape-of-Motiongeometric 4D-GS

self-PSNR 60.3 dB · GT-PSNR (full-frame) 18.6 dB · GT-LPIPS 0.424

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 69.7 · photo ↑ 0.0 (EPE 1.66px) · motion 10.94 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

GEN3C3D-cache

self-PSNR 17.7 dB · GT-PSNR (full-frame) 16.4 dB · GT-LPIPS 0.318

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 36.1 · photo ↑ 0.0 (EPE 1.93px) · motion 6.18 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

MatrixGame-3learned

self-PSNR 14.0 dB · GT-PSNR (full-frame) 10.9 dB · GT-LPIPS 0.818

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Ours VACE-14Bwarp+inpaint

self-PSNR 13.8 dB · GT-PSNR (full-frame) 14.7 dB · GT-LPIPS 0.388

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 56.5 · photo ↑ 42.4 (EPE 0.69px) · motion 6.71 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Ours VACE-14B (revisit-FT)control-only FT

self-PSNR 13.0 dB · GT-PSNR (full-frame) 13.2 dB · GT-LPIPS 0.436

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 57.6 · photo ↑ 44.7 (EPE 0.66px) · motion 6.51 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Ours VACE-14B (revisit-FT, DiT+VACE-full)DiT+VACE full FT

self-PSNR 12.9 dB · GT-PSNR (full-frame) 14.1 dB · GT-LPIPS 0.471

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 38.1 · photo ↑ 11.3 (EPE 1.06px) · motion 6.77 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

SpMemspatial-mem

self-PSNR 12.2 dB · GT-PSNR (full-frame) 11.4 dB · GT-LPIPS 0.734

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 52.9 (EPE 0.56px) · motion 3.99 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Lyra-2learned

self-PSNR 11.7 dB · GT-PSNR (full-frame) 10.2 dB · GT-LPIPS 0.677

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Voyagerworld-model

metrics pending

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

FrameCraftersingle-shot

self-PSNR 16.8 dB · GT-PSNR (full-frame) 16.3 dB · GT-LPIPS 0.371

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

ViewCraftersingle-shot

self-PSNR 14.4 dB · GT-PSNR (full-frame) 13.0 dB · GT-LPIPS 0.499

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 56.6 · photo ↑ 0.0 (EPE 1.59px) · motion 7.36 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

NVS-Solversingle-shot

self-PSNR 11.6 dB · GT-PSNR (full-frame) 11.0 dB · GT-LPIPS 0.744

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Geometric-memory methods (Shape-of-Motion) reproduce the revisited pose near-perfectly (self-PSNR often >50 dB), while learned / single-shot methods drift between visits (self-PSNR ~12–19 dB). Numbers are read directly from the revisit metrics (not recomputed). Dynamic-region PSNR scores the mover pixels only (spatialvid dyn-masks / kubric segmentation, background & >35%-area static planes excluded) — the full-frame GT-PSNR is inflated for warp/copy methods by static-background copying. DL3DV has no masks → n/a.

WorldScore consistency ⟳ is reference-free — computed from each method's own generated frames only (no GT poses/depth/frames; the requested camera path is NOT checked). 3D↑ = DROID dense-BA reprojection self-consistency (1−e/1.0719, ×100) — explicit-geometry methods (GEN3C 3D-cache, warp/4D-cache) score highest; a real DROID tracking-loss is recorded as an honest 0. photo↑ = SEA-RAFT forward/backward flow-cycle stability (1−EPE/1.192, ×100); this bound is tuned for smooth single-scene static clips, so under the large inter-frame motion of these revisit clips it saturates near 0 — the raw EPE (px) is shown alongside as the discriminative number. motion = median flow (px), the guard-rail: a frozen/degenerate clip fakes ~100 consistency at ~0 motion, so a high consistency next to near-zero motion is a frozen-cheat, not real memory. Scored on the shown clips for {base VACE-14B, revisit-FT, DiT+VACE-full, 4D-cache} + {GEN3C, Shape-of-Motion, SpMem, ViewCrafter}; other methods show “not scored”.

Same-pose return frames revisit fidelity grouped by pose: each box leads with the OBSERVED GT for that pose (the reference target), then that method's Visit 1 / 2 / 3 of the SAME pose. self-PSNR = cross-visit consistency, GT-PSNR = match to the real observed frame.

Shape-of-Motiongeometric 4D-GS

self-PSNR 60.6 dB · GT-PSNR (full-frame) 21.3 dB · GT-LPIPS 0.487

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 23.89px) · motion 35.92 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

GEN3C3D-cache

self-PSNR 18.6 dB · GT-PSNR (full-frame) 19.8 dB · GT-LPIPS 0.386

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 4.07px) · motion 12.21 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

SpMemspatial-mem

self-PSNR 18.2 dB · GT-PSNR (full-frame) 13.5 dB · GT-LPIPS 0.742

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 4.70px) · motion 24.97 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Ours VACE-14Bwarp+inpaint

self-PSNR 18.0 dB · GT-PSNR (full-frame) 17.2 dB · GT-LPIPS 0.620

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 11.55px) · motion 38.22 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Lyra-2learned

self-PSNR 17.4 dB · GT-PSNR (full-frame) 15.4 dB · GT-LPIPS 0.690

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Ours VACE-14B (revisit-FT)control-only FT

self-PSNR 16.8 dB · GT-PSNR (full-frame) 15.5 dB · GT-LPIPS 0.627

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 13.52px) · motion 34.75 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Ours VACE-14B (revisit-FT, DiT+VACE-full)DiT+VACE full FT

self-PSNR 16.7 dB · GT-PSNR (full-frame) 17.1 dB · GT-LPIPS 0.590

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 23.49px) · motion 40.88 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

MatrixGame-3learned

self-PSNR 11.9 dB · GT-PSNR (full-frame) 12.9 dB · GT-LPIPS 0.732

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Voyagerworld-model

metrics pending

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

FrameCraftersingle-shot

self-PSNR 21.8 dB · GT-PSNR (full-frame) 20.6 dB · GT-LPIPS 0.285

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

ViewCraftersingle-shot

self-PSNR 17.0 dB · GT-PSNR (full-frame) 14.7 dB · GT-LPIPS 0.701

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 18.71px) · motion 34.11 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

NVS-Solversingle-shot

self-PSNR 10.8 dB · GT-PSNR (full-frame) 11.5 dB · GT-LPIPS 0.730

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Geometric-memory methods (Shape-of-Motion) reproduce the revisited pose near-perfectly (self-PSNR often >50 dB), while learned / single-shot methods drift between visits (self-PSNR ~12–19 dB). Numbers are read directly from the revisit metrics (not recomputed). Dynamic-region PSNR scores the mover pixels only (spatialvid dyn-masks / kubric segmentation, background & >35%-area static planes excluded) — the full-frame GT-PSNR is inflated for warp/copy methods by static-background copying. DL3DV has no masks → n/a.

WorldScore consistency ⟳ is reference-free — computed from each method's own generated frames only (no GT poses/depth/frames; the requested camera path is NOT checked). 3D↑ = DROID dense-BA reprojection self-consistency (1−e/1.0719, ×100) — explicit-geometry methods (GEN3C 3D-cache, warp/4D-cache) score highest; a real DROID tracking-loss is recorded as an honest 0. photo↑ = SEA-RAFT forward/backward flow-cycle stability (1−EPE/1.192, ×100); this bound is tuned for smooth single-scene static clips, so under the large inter-frame motion of these revisit clips it saturates near 0 — the raw EPE (px) is shown alongside as the discriminative number. motion = median flow (px), the guard-rail: a frozen/degenerate clip fakes ~100 consistency at ~0 motion, so a high consistency next to near-zero motion is a frozen-cheat, not real memory. Scored on the shown clips for {base VACE-14B, revisit-FT, DiT+VACE-full, 4D-cache} + {GEN3C, Shape-of-Motion, SpMem, ViewCrafter}; other methods show “not scored”.

Same-pose return frames revisit fidelity grouped by pose: each box leads with the OBSERVED GT for that pose (the reference target), then that method's Visit 1 / 2 / 3 of the SAME pose. self-PSNR = cross-visit consistency, GT-PSNR = match to the real observed frame.

Shape-of-Motiongeometric 4D-GS

self-PSNR 62.1 dB · GT-PSNR (full-frame) 25.3 dB · GT-LPIPS 0.441

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 24.36px) · motion 49.66 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Ours VACE-14B (revisit-FT)control-only FT

self-PSNR 19.1 dB · GT-PSNR (full-frame) 19.7 dB · GT-LPIPS 0.611

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 20.01px) · motion 56.68 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Lyra-2learned

self-PSNR 19.1 dB · GT-PSNR (full-frame) 17.3 dB · GT-LPIPS 0.607

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Ours VACE-14B (revisit-FT, DiT+VACE-full)DiT+VACE full FT

self-PSNR 18.8 dB · GT-PSNR (full-frame) 18.6 dB · GT-LPIPS 0.602

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 27.60px) · motion 48.79 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Ours VACE-14Bwarp+inpaint

self-PSNR 18.2 dB · GT-PSNR (full-frame) 19.8 dB · GT-LPIPS 0.669

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 11.26px) · motion 42.85 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

SpMemspatial-mem

self-PSNR 18.0 dB · GT-PSNR (full-frame) 15.5 dB · GT-LPIPS 0.718

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 24.0 · photo ↑ 0.0 (EPE 5.52px) · motion 30.27 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

MatrixGame-3learned

self-PSNR 17.9 dB · GT-PSNR (full-frame) 12.7 dB · GT-LPIPS 0.685

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

GEN3C3D-cache

self-PSNR 17.8 dB · GT-PSNR (full-frame) 17.6 dB · GT-LPIPS 0.533

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 57.9 · photo ↑ 0.0 (EPE 5.07px) · motion 22.86 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

Voyagerworld-model

metrics pending

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

ViewCraftersingle-shot

self-PSNR 23.3 dB · GT-PSNR (full-frame) 19.0 dB · GT-LPIPS 0.456

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free 3D ↑ 0.0 · photo ↑ 0.0 (EPE 6.90px) · motion 36.82 px

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

FrameCraftersingle-shot

self-PSNR 17.6 dB · GT-PSNR (full-frame) 18.8 dB · GT-LPIPS 0.413

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176

NVS-Solversingle-shot

self-PSNR 15.2 dB · GT-PSNR (full-frame) 15.9 dB · GT-LPIPS 0.674

dynamic-region PSNR n/a (no mover mask for this benchmark)

WorldScore consistency ⟳ ref-free not scored

Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1).

Pose A first observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose A · Observed GTfirst observed frame

Visit 1gen · f73

Visit 2gen · f121

Visit 3gen · f169

Pose B mid observed frame — revisited 3×; these 3 visits should reproduce the Observed GT on the left

Pose B · Observed GTmid observed frame

Visit 1gen · f80

Visit 2gen · f128

Visit 3gen · f176