Anonymous submission. Within each figure every column/row starts from the same initial noise. Every section shows the same 18 prompts of the 100-prompt test set, numbered 1 to 18 in the same order (selected examples; the tables cover all 100 prompts). In every table Δ = PTD − other (a positive value favors PTD); * = significant.
K=1 (learned curve) vs. K=0 (straight-chord queries), trained identically: same prompts, seed, teacher, query budget and updates; only the curve that places the teacher queries differs. Columns: 50-step teacher | K=0 | K=1.
VBench (946 prompts × 5 videos):
| total | dynamic degree | temporal flickering | |
|---|---|---|---|
| K=0 (straight chord) | 83.71 | 74.7 | 99.21 |
| K=1 (learned curve) | 83.84 | 76.7 | 98.75 |
Blinded pairwise judge (100 prompts × 3 seeds; Δ = learned curve − K=0):
| dynamic | static | naturalness | diversity | |
|---|---|---|---|---|
| curve − K=0 | +.05 | +.10* | +.21* | −.00 |
Reading. K=0 produces red and blue color artifacts that the learned curve does not; the judge rates the curve more natural (+.21) with better per-frame quality (+.10).
Columns: teacher (50 steps) · K=0 (straight chord) · K=1 (learned curve)
K=0 is trained with the paper's H3 recipe (150 updates; same data, teacher, adapters and final-hop march) but without the curvature head, so the two teacher queries per sample lie on the straight chord; it is compared with the paper's model (learned curve). Columns: 50-step teacher | paper model | K=0.
Blinded pairwise judge (100 prompts × 3 seeds; Δ = paper model − K=0):
| paper model − K=0 | dynamic | static | naturalness | diversity |
|---|---|---|---|---|
| all 100 prompts | +.09* | +.03 | +.03 | −.07 |
| 50 action prompts | +.14* | +.01 | −.01 | −.12* |
Reading. The paper model (learned curve) is more dynamic than K=0 (+.09; +.14 on the action prompts), and K=0 adds fine-scale frame-to-frame flicker beyond the teacher (finest-band power 120% of the teacher for K=0, 105% for the paper model); overall motion is unchanged.
Columns: teacher (50 steps) · paper model (learned curve) · K=0 (no curvature head)
Each baseline is trained with PTD's Wan recipe (64 updates, batch 256, the same teacher, 4-step inference) with one component changed, and compared with PTD@64. Rows: PTD, Euler endpoint regression, LMD | PTD, MeanFlow-style, MeanFlow-style with 75% flow-matching samples. Judge Δ = PTD − baseline:
| dynamic | static | naturalness | diversity | |
|---|---|---|---|---|
| PTD − Euler endpoint regression | +.64* | +1.04* | +.73* | +.05 |
| PTD − LMD (two-time map) | +.32* | +.08* | −.11* | +.86* |
| PTD − MeanFlow-style | +1.04* | +2.00* | +1.91* | +.69* |
| PTD − MeanFlow-style, 75% r = t | +.73* | +1.43* | +1.34* | +.50* |
Top: PTD · Euler endpoint · LMD — Bottom: PTD · MeanFlow-style · MeanFlow-style 75%
LightX2V's public four-step Wan2.1-14B student (distribution matching, its own sampler) vs. PTD; seeds 0–2 per row. Judge Δ = PTD − DMD:
| dynamic | static | naturalness | diversity | |
|---|---|---|---|---|
| PTD − DMD | +.11 | −.07 | +.40* | +.83* |
Top: PTD, seeds 0–2 · Bottom: DMD, seeds 0–2