PermaVid revisit benchmark · span 0.33 (slow / small-motion)

Span 0.33 revisit — video, degradation, and same-pose model comparison

The slow-sweep setting (span_frac = 0.33): the camera moves gently and revisits observed viewpoints. Each scene has every method as a playable video (progress bar + 0.25–1× speed), a multi_visit degradation curve (per-frame PSNR with revisit frames shaded), and a same-camera-pose comparison ranking all models at the returning pose.

Play all fires every method in a scene in sync (kept frame-aligned to the GT clock). Drag the GT progress bar to scrub the whole scene together; the speed buttons apply to all at once.

🎬 The condition (what every method is given): the first 36 observed frames (the blue CONDITION clip in each scene) + the full 180-pose camera trajectory (intrinsics K + per-frame camera-to-world). The task: generate the remaining 144 frames following that exact path — which loops back to observed viewpoints, so it tests whether the model remembers the scene on return.

speed
Loop-closure metrics @ span 0.33 — perceptual fidelity at the revisit frames (vs the real GT return). Ranked by DreamSim ↓.
#methodDreamSim↓LPIPS (LCE)↓revisit PSNR↑
1FrameCrafter0.2020.39815.00
2NVS-Solver0.2670.62612.27
3Lyra 2.00.2850.57813.30
4GEN3C0.2950.56814.30
5FlexWorld0.3060.57013.29
6VACE-14B (finetuned)0.3190.60312.40
7MatrixGame-30.3690.65611.18
8ViewCrafter0.4220.60811.35
9SpMem0.4220.65712.86
10Voyager0.4280.72111.91
11Shape-of-Motion0.5160.77910.56
12ViewExtrapolator0.5170.70911.37
13TrajectoryCrafter0.6220.76111.10
Map_ChemicalPlant_113 methods · 16 revisit frames · span 0.33speed
CONDITION · 36 observed frames (input context)
GROUND TRUTH (real render)
VACE-14B (finetuned)
GEN3C
TrajectoryCrafter
FrameCrafter
Lyra 2.0
SpMem
MatrixGame-3
NVS-Solver
FlexWorld
ViewExtrapolator
ViewCrafter
Voyager
Shape-of-Motion
mid revisit — frame 132: camera returns to observed pose 26. Same camera pose → models ranked by PSNR vs the GT return.
GT return
real render @ frame 132
1. Shape-of-Motion
13.4 dB
2. FlexWorld
13.0 dB
3. GEN3C
13.0 dB
4. VACE-14B (finetuned)
11.9 dB
5. Lyra 2.0
11.8 dB
6. TrajectoryCrafter
10.7 dB
7. SpMem
10.1 dB
8. FrameCrafter
10.1 dB
9. ViewExtrapolator
9.4 dB
10. MatrixGame-3
9.3 dB
11. NVS-Solver
9.2 dB
12. ViewCrafter
8.5 dB
13. Voyager
6.2 dB
deep revisit (hardest) — frame 152: camera returns to observed pose 35. Same camera pose → models ranked by PSNR vs the GT return.
GT return
real render @ frame 152
1. GEN3C
12.7 dB
2. FlexWorld
10.9 dB
3. Shape-of-Motion
10.5 dB
4. VACE-14B (finetuned)
10.0 dB
5. TrajectoryCrafter
9.8 dB
6. Lyra 2.0
9.8 dB
7. FrameCrafter
8.7 dB
8. SpMem
8.4 dB
9. NVS-Solver
7.4 dB
10. ViewCrafter
7.4 dB
11. ViewExtrapolator
7.3 dB
12. MatrixGame-3
7.3 dB
13. Voyager
6.8 dB
Medieval_Nighttime12 methods · 1 revisit frames · span 0.33speed
CONDITION · 36 observed frames (input context)
GROUND TRUTH (real render)
VACE-14B (finetuned)
GEN3C
TrajectoryCrafter
FrameCrafter
Lyra 2.0
SpMem
MatrixGame-3
NVS-Solver
FlexWorld
ViewExtrapolator
ViewCrafter
Voyager
This scene barely revisits at span 0.33 (1 revisit frame) — video + curve only.
ModularVictorianCity12 methods · 71 revisit frames · span 0.33speed
CONDITION · 36 observed frames (input context)
GROUND TRUTH (real render)
VACE-14B (finetuned)
GEN3C
TrajectoryCrafter
FrameCrafter
Lyra 2.0
SpMem
MatrixGame-3
NVS-Solver
FlexWorld
ViewExtrapolator
ViewCrafter
Voyager
mid revisit — frame 71: camera returns to observed pose 26. Same camera pose → models ranked by PSNR vs the GT return.
GT return
real render @ frame 71
1. FrameCrafter
18.1 dB
2. Lyra 2.0
17.3 dB
3. TrajectoryCrafter
16.8 dB
4. MatrixGame-3
16.1 dB
5. NVS-Solver
16.1 dB
6. Voyager
15.8 dB
7. FlexWorld
15.7 dB
8. VACE-14B (finetuned)
15.7 dB
9. GEN3C
15.6 dB
10. ViewExtrapolator
14.4 dB
11. SpMem
13.0 dB
12. ViewCrafter
11.8 dB
deep revisit (hardest) — frame 148: camera returns to observed pose 14. Same camera pose → models ranked by PSNR vs the GT return.
GT return
real render @ frame 148
1. GEN3C
13.2 dB
2. TrajectoryCrafter
12.3 dB
3. Lyra 2.0
12.2 dB
4. MatrixGame-3
12.0 dB
5. Voyager
11.9 dB
6. FlexWorld
11.7 dB
7. SpMem
11.4 dB
8. FrameCrafter
10.8 dB
9. VACE-14B (finetuned)
10.7 dB
10. NVS-Solver
10.5 dB
11. ViewExtrapolator
10.4 dB
12. ViewCrafter
10.1 dB
Science_Fiction_valley_town12 methods · 20 revisit frames · span 0.33speed
CONDITION · 36 observed frames (input context)
GROUND TRUTH (real render)
VACE-14B (finetuned)
GEN3C
TrajectoryCrafter
FrameCrafter
Lyra 2.0
SpMem
MatrixGame-3
NVS-Solver
FlexWorld
ViewExtrapolator
ViewCrafter
Voyager
mid revisit — frame 46: camera returns to observed pose 35. Same camera pose → models ranked by PSNR vs the GT return.
GT return
real render @ frame 46
1. FrameCrafter
17.3 dB
2. FlexWorld
16.7 dB
3. GEN3C
15.5 dB
4. ViewCrafter
15.2 dB
5. Lyra 2.0
14.3 dB
6. SpMem
14.0 dB
7. Voyager
13.6 dB
8. NVS-Solver
12.9 dB
9. ViewExtrapolator
12.5 dB
10. MatrixGame-3
12.1 dB
11. VACE-14B (finetuned)
11.5 dB
12. TrajectoryCrafter
8.6 dB
deep revisit (hardest) — frame 55: camera returns to observed pose 35. Same camera pose → models ranked by PSNR vs the GT return.
GT return
real render @ frame 55
1. FrameCrafter
17.6 dB
2. GEN3C
16.3 dB
3. FlexWorld
15.6 dB
4. ViewCrafter
14.0 dB
5. Lyra 2.0
13.7 dB
6. SpMem
13.7 dB
7. Voyager
13.5 dB
8. NVS-Solver
12.5 dB
9. MatrixGame-3
12.4 dB
10. VACE-14B (finetuned)
11.7 dB
11. ViewExtrapolator
10.3 dB
12. TrajectoryCrafter
9.1 dB
Western_Garden11 methods · 15 revisit frames · span 0.33speed
CONDITION · 36 observed frames (input context)
GROUND TRUTH (real render)
VACE-14B (finetuned)
GEN3C
TrajectoryCrafter
FrameCrafter
Lyra 2.0
MatrixGame-3
NVS-Solver
FlexWorld
ViewExtrapolator
ViewCrafter
Voyager
mid revisit — frame 79: camera returns to observed pose 30. Same camera pose → models ranked by PSNR vs the GT return.
GT return
real render @ frame 79
1. FlexWorld
13.0 dB
2. GEN3C
11.2 dB
3. VACE-14B (finetuned)
11.1 dB
4. FrameCrafter
10.9 dB
5. Voyager
9.7 dB
6. ViewCrafter
9.6 dB
7. TrajectoryCrafter
9.5 dB
8. ViewExtrapolator
9.4 dB
9. Lyra 2.0
8.6 dB
10. NVS-Solver
8.5 dB
11. MatrixGame-3
8.4 dB
deep revisit (hardest) — frame 146: camera returns to observed pose 14. Same camera pose → models ranked by PSNR vs the GT return.
GT return
real render @ frame 146
1. Lyra 2.0
12.9 dB
2. FlexWorld
11.5 dB
3. TrajectoryCrafter
10.5 dB
4. VACE-14B (finetuned)
10.4 dB
5. GEN3C
10.3 dB
6. FrameCrafter
9.7 dB
7. Voyager
9.1 dB
8. ViewExtrapolator
8.8 dB
9. ViewCrafter
8.3 dB
10. MatrixGame-3
7.9 dB
11. NVS-Solver
7.8 dB