acceptodds
Under review as a conference paper at ICLR 2027

Parametric Trajectory Distillation for Few-Step Video Generation

Abstract

Video diffusion and flow models require many sequential evaluations, making generation computationally expensive. Few-step distillation reduces this cost but poses a capacity allocation problem: a student must match the teacher's iterative generation with far less sequential computation. Existing trajectory methods ask the student to reproduce teacher transitions that are highly curved at high noise, which can exceed its capacity and degrade fine detail. We introduce Parametric Trajectory Distillation (PTD), which lets the student parameterize teacher trajectory segments as polynomials and learn from teacher guidance along its own predicted path. PTD is designed to let the learned curvature adapt to the backbone's predictive capacity, preserving motion and diversity. The curvature head is used only in training; inference keeps the original backbone architecture. On Wan2.1-14B, four-step PTD sets a new state of the art for trajectory distillation, significantly improving dynamic quality and naturalness over PDD, the best-performing trajectory-only method on this model, under the same training setting. On the 33B audio-video MiniMax-H3, LoRA-trained PTD significantly improves diversity and naturalness over the state-of-the-art LightX2V Turbo. Blinded human votes give PTD 55.1% and 63.4% preference shares against PDD and LightX2V Turbo. Anonymous project page: https://ptd-website.github.io/.

Primary area:
generative models
Keywords:
Trajectory Distillation, Parametric Trajectory Modeling, Few-Step Video Generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_rUYtOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Parametric Trajectory Distillation (PTD), which represents student trajectories using a displacement prediction and polynomial shape coefficients. Training matches velocities along the predicted trajectory to teacher velocities, while inference uses only the displacement prediction. The paper evaluates the method on Wan2.1-14B against PDD and on MiniMax-H3 against LightX2V Turbo.

Critical strengths

The parameterization enables inexpensive intermediate trajectory supervision without network Jacobian-vector products. The H3 results show promising improvements in diversity and naturalness while maintaining comparable visual quality. Establishing this advantage over DMD across broader settings would be a highly valuable contribution.

Critical weaknesses
  1. The controlled trajectory-distillation comparison relies primarily on PDD, another ICLR 2027 submission still under review. Direct comparisons with established alternatives, including LMD and MeanFlow-based distillation, are missing. The multi-step Euler endpoint-supervision baseline is evaluated only through limited qualitative examples, leaving the practical advantage over simpler approaches unclear.

  2. The paper does not sufficiently explain or validate why retaining trajectory constraints produces a better quality–diversity trade-off than distribution matching. The DMD comparison is limited to one released checkpoint on H3, whose PTD recipe also includes final-hop teacher integration. Broader controlled experiments are needed to establish the claimed advantage.

Questions for authors
  1. How does PTD compare quantitatively with multi-step Euler endpoint regression, LMD, and MeanFlow-based distillation under comparable training and inference budgets?

  2. What evidence shows that the gains over DMD persist across backbones and training settings, and arise from the proposed trajectory supervision?

3 comments
AuthorsOpenReview ↗

Reply to Reviewer rUYt (1/2)

Thank you for the careful review and the concrete questions. Every comparison in our two replies uses the paper's recipes and judging protocol (Sec. 4.1, App. C: four Astra axes on 100 prompts × 3 seeds); Δ = PTD − other, so a positive value favors PTD.

W1 / Q1: established alternatives at comparable budgets. PTD beats every alternative we trained with its recipe on every quality axis except LMD's naturalness, and it is the only one that keeps the seed diversity. The paper reproduces PDD, the strongest trajectory-only method reported on Wan2.1-14B, under PTD's exact setting. We add three more on Wan2.1-14B, each trained with PTD's data, updates, batch, teacher and 4-NFE inference and changing only the training target or the map (videos):

Δ = PTD@64 − alternative (* significant)teacher velocities / samples / updatedynamicstaticnatural.diversityVBenchdyn. degreeseed div. (V-JEPA 2)
PTD@64 (reference)211884.5776.1.122
Euler endpoint regression (8-step teacher march, every hop)8287+.64*+1.04*+.73*+.0580.9356.4.115
LMD (conventional two-time map, same 2 target times)2342+.32*+.08*−.11*+.86*83.1848.3.074
MeanFlow-style (two-time map, Eulerian identity)1193+1.04*+2.00*+1.91*+.69*69.9837.5.102
MeanFlow-style, MeanFlow's default sampling (75% equal-time samples)1192+.73*+1.43*+1.34*+.50*
  • Euler endpoint regression: the student regresses the mean velocity of an 8-step guided teacher march from the carried state, on every hop (the quantitative version of Fig. 6). It costs 2.4× more per update and loses on every quality axis, the detail loss that Fig. 6 shows; diversity is unchanged.
  • LMD: the conventional two-time Lagrangian map at the same two target times. PTD is more dynamic, sharper and far more diverse; LMD is more natural by .11, at 2.9× the cost per update, with 37% of the teacher's motion energy (PTD 118%) and a seed diversity at the DMD student's level.
  • MeanFlow-based distillation: the same two-time average-velocity network, trained with MeanFlow's Eulerian identity (the average velocity equals the teacher velocity plus the time interval times the map's time derivative), a stop-gradient target, a finite-difference derivative along the teacher flow and MeanFlow's adaptive weighting. Its final-hop map diverged between 32 and 64 updates and left visible noise in the samples. With MeanFlow's default sampling it improves and stays far behind PTD on all four axes.
AuthorsOpenReview ↗

Reply to Reviewer rUYt (2/2)

W2 / Q2: DMD across backbones; final-hop integration. (Same protocol as in part 1/2; Δ = PTD − other.)

Second backbone. Against LightX2V's released four-step Wan2.1-T2V-14B DMD student, PTD@64 repeats the H3 pattern (first row of the next table): static and dynamic quality on par, naturalness and diversity higher (videos). VBench does not separate the two on Wan (PTD 82.92 to 84.57 across checkpoints); seed diversity confirms the gap.

Wan2.1-14B (* significant)dynamicstaticnaturalnessdiversityVBenchseed div. (V-JEPA 2)
PTD@64 − LightX2V 4-step DMD student+.11−.07+.40*+.83*84.57 vs. 83.97.122 vs. .078

Final-hop integration and the adapter budget. On H3 we retrained PTD without the final-hop march (two teacher queries on every hop) and with one shared rank-128 LoRA for all steps (Turbo's adapter budget), otherwise with the paper's recipe:

MiniMax-H3, Δ vs. LightX2V Turbo v1.1 (* significant)dynamicstaticnaturalnessdiversity
Paper model (re-judged with the new rows)+.11−.02+.14*+1.30*
No final march (2 queries on all hops)−.39*−.45*−.61*+1.36*
One shared rank-128 LoRA (Turbo's budget)+.04−.08*+.09*+1.50*

The final-hop march supplies the last hop's fine detail (Sec. 3.2): without it PTD falls .56, .50 and .83 below its own paper model on dynamic quality, static quality and naturalness, while diversity stays the same (+.06), so the march is not the source of PTD's diversity. With Turbo's adapter budget, PTD stays more natural and far more diverse than Turbo.

Why trajectory supervision keeps diversity. Regressing the teacher velocity along the student's own path with a single-time network keeps seed diversity on both backbones (PTD − DMD: +1.30 on H3, +.83 on Wan), consistent with per-noise teacher supervision preserving the teacher's noise-to-sample coupling; reverse-KL distribution matching is mode-seeking. The two-time LMD map, trained on the same kind of trajectories, loses that diversity (PTD − LMD: +.86), which points to PTD's single-time deployed map, with its target time confined to the training-only head, as the responsible design choice. The revision states this mechanism with the LMD control.

Thank you for the detailed responses and additional experiments. The comparisons with established baselines, the additional DMD comparison on Wan2.1-14B, and the ablations on final-hop integration and shared LoRA address my concerns. Based on the rebuttal, I will raise my score from 2 to 3.

Reviewer_6QTkOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Parametric Trajectory Distillation (PTD) for four-step video generation. At each denoising step, the student predicts a mean velocity and a polynomial curve, and teacher velocities sampled along this curve are used as training targets. At inference, the curve head is removed and only the mean velocity is used. Experiments on Wan2.1-14B and MiniMax-H3 show gains in motion quality and diversity over the compared baselines.

Critical strengths
  1. The method separates the endpoint update from the path shape inside each step. The polynomial curve is only used to query the teacher during training, while inference uses just the mean velocity. This lets the student drop the curve head and keep the original backbone and output structure at test time.
  2. On MiniMax-H3, diversity improves from 1.55 to 2.81 over LightX2V Turbo, and human raters prefer PTD with a 63.4% share. The experiments also cover both full fine-tuning (Wan2.1-14B) and LoRA (H3), showing that the training recipe works in both settings.
Critical weaknesses
  1. The ablation study is missing a K=0 baseline. A straight-path student would still query the teacher along its own predicted path while removing the learned curvature terms. Comparing K=0 with the default K=8 model under the same training and query budgets is needed to show whether the curvature branch actually improves final video quality.
  2. The link between learned curvature and final video quality is not directly tested. Figures 5 and 8 measure curvature alignment, fitting, and prediction, while Figure 6 only shows qualitative examples of different supervision choices. These results do not show how changes in learned curvature relate to endpoint error and final video quality. Reporting these metrics together across different K values or training checkpoints would directly test the paper's main claim.
  3. The MiniMax-H3 experiment trains a separate LoRA for each step without an ablation. On H3, PTD uses a separate rank-128 LoRA adapter for each of the four steps. Some works(https://arxiv.org/pdf/2410.06664?, https://arxiv.org/pdf/2506.03123) have shown that training step-specific LoRAs can improve few-step distillation on its own. Without an ablation using a single shared LoRA, it is unclear whether the gains on H3 come from PTD itself or from the per-step LoRAThinking Process.
Questions for authors

Please see the weekness.

3 comments
AuthorsOpenReview ↗

Reply to Reviewer 6QTk (1/2)

Thank you for the detailed review. We ran all three requested experiments.

W1: K=0 at the same training and query budgets. The learned curve helps on both backbones we tested (Δ = curve − K=0; * significant).

MiniMax-H3, end to end. K=0 trains with the paper's H3 recipe (150 updates) without the curvature head, so the teacher queries lie on the chord; the comparison is the paper model (K=8):

paper model − K=0dynamicstaticnaturalnessdiversity
all 100 prompts+.09*+.03+.03−.07
50 action prompts+.14*+.01−.01−.12*

K=0 also adds fine-scale flicker that the 12-frame judge cannot see: on the full videos its finest-band power is 120% of the teacher's, the paper model's 105%. Side-by-side videos: teacher, paper model, K=0.

Wan2.1-1.3B. K=1 (learned curve) against K=0 (straight chord) with identical prompts, seed, teacher, query budget and updates:

K=1 − K=0dynamicstaticnaturalnessdiversity
Astra Δ+.05+.10*+.21*.00
VBenchdynamic degree
K=083.7174.7
K=183.8476.7

The curve makes the student more natural and sharper; K=0 shows red and blue color artifacts that the curve does not. Side-by-side videos: teacher, K=0, K=1.

The revision reports both comparisons in Sec. 4.4 and App. D.5 (Table 15).

AuthorsOpenReview ↗

Reply to Reviewer 6QTk (2/2)

W2: learned curvature, endpoint error and video quality. With the probe of Fig. 5 / App. D (the bend magnitude ratio, the teacher-aligned bend projection, and the endpoint error per hop, anchored to the teacher), across the paper's PTD-14B training run at every saved checkpoint:

updates326496128160192224
hop-0 bend ratio / teacher-aligned bend.05 / .010.08 / .014.11 / .021.14 / .033.18 / .040.19 / .046.20 / .054
hop-3 teacher-aligned bend.011.022.031.038.045.053.061
endpoint err. hop 0 / hop 3.62 / .114.72 / .106.75 / .106.66 / .128.74 / .122.70 / .129.70 / .121
VBench84.4684.5783.2984.0082.9283.9183.86

Two readings. The student learns a teacher-aligned bend that grows steadily with training, as in Fig. 5. Endpoint error does not predict video quality: a student trained with the same recipe to regress teacher-integrated endpoints (an 8-step teacher march from the carried state, on every hop) has the lowest endpoint error of all the Wan2.1-14B variants we trained (hop 0: .46 vs. .72 for PTD@64) and a static quality 1.04 below PTD@64, and VBench across the checkpoints moves with its dynamic-degree dimension (38.9 to 85.3) while the endpoint error stays nearly flat. The curve's effect on quality comes from where it places the teacher queries (W1). This series is Table 14 in App. D.3 of the revision.

W3: per-step LoRA confound on H3. Thank you for the pointers to DeMe and DCM, which we now cite. LightX2V Turbo itself uses one rank-128 LoRA for all four steps (692M parameters). We trained PTD with one shared rank-128 adapter set for all hops (hops then differ only through the timestep), otherwise with the paper's H3 recipe, and judged it under the paper's protocol against Turbo and against the paper's per-hop model:

MiniMax-H3 (* significant)adaptersdynamicstaticnaturalnessdiversity
Paper model − Turbo (re-judged with the new rows)4 × rank-128+.11−.02+.14*+1.30*
Shared-LoRA PTD − Turbo1 × rank-128+.04−.08*+.09*+1.50*
Shared-LoRA PTD − paper model1 vs. 4 × rank-128−.04−.05*−.03+.12*

With exactly Turbo's adapter budget, PTD stays more natural and far more diverse than Turbo, with dynamic quality on par and static quality .08 lower. Against the paper's per-hop model, the shared adapter differs by at most .05 on any quality axis and is more diverse. PTD's advantages over DMD hold with a single shared adapter.

I thank the authors for the thorough response and for running addtional ablations, these results have already addressed my concerns. I am raising my score to 3 (Weak acceptance).

Reviewer_yYv8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Parametric Trajectory Distillation (PTD) for few-step video generation. A student predicts an interval-average velocity and polynomial curvature coefficients in one forward pass. Teacher velocities queried along the student-predicted curve supervise its analytic tangent. The curvature head is removed at inference, retaining four-step sampling. Experiments on several models show improvements over previous distillation methods.

Critical strengths

-The separation of endpoint displacement from interior curvature is elegant. It allows the curvature head to guide training without adding inference overhead.

-The experiments cover two large different video backbones, demonstrating practical applicability.

-The appendices provide many implementation details.

Critical weaknesses

I generally think the proposed parametric method has good practical value in application. I have some minor concerns regarding the novelty and experiments.

-The method novelty appears incremental. As described in Appendix A.4, the objective specializes existing Lagrangian map distillation using a polynomial parameterization. Comparisons with a conventional Lagrangian-map baseline or pi-Flow would better demonstrate the benefits of the method.

-Although the paper studies different coefficient counts through offline fitting and frozen-feature probes, it does not report an end-to-end K=0 straight-path baseline. This comparison would clarify whether learned curvature improves generation quality beyond student-path supervision alone.

Questions for authors

I'd like to see the response on novelty and additional experiments.

1 comment

Thank you for the positive assessment and the two concrete requests. Every comparison below uses the paper's recipes and judging protocol (Sec. 4.1, App. C: four Astra axes on 100 prompts × 3 seeds); Δ = PTD − other, so a positive value favors PTD.

W1 / Q1: novelty relative to LMD; LMD or π-Flow baseline. PTD's contribution is the parameterization that makes a restricted LMD loss (App. A.4: grid start times, rollout start states, a polynomial target-time basis, two stratified target times per sample) practical for large video backbones and changes what is deployed:

  1. Single-time interface, removable head. The target time enters only through a training-only polynomial head, and the endpoint depends only on the single-time network (App. A.4), so inference runs the unchanged backbone with 4 NFE. A conventional LMD map keeps its target-time input at deployment.
  2. Analytic tangent. All target-time queries share one student pass and ordinary backpropagation; LMD needs the derivative of its map with respect to the target time, by a Jacobian-vector product (about 1.5 forward passes, Boffi et al. 2025) or by extra passes.
  3. Student-path queries with a capacity-limited curve, tested directly below: against regressing teacher-integrated endpoints (Euler row of the table) and against the straight chord (W2).

New: LMD trained with PTD's recipe. Each row trains Wan2.1-14B with the paper's Wan recipe (App. B.2: batch 256, 64 updates, same teacher, data and target times) and changes only the map; the reference is PTD@64. LMD here is the conventional two-time network that reads the start state, the start time and the target time and predicts the average velocity between the two times; we train it with the stop-gradient Lagrangian residual of App. A.4 and take the target-time derivative by a central difference.

Δ = PTD@64 − alternative (* significant)s / updatedynamicstaticnatural.diversityVBenchseed div. (V-JEPA 2)
PTD@64 (reference)11884.57.122
LMD (two-time map, same 2 target times)342+.32*+.08*−.11*+.86*83.18.074
Euler endpoint regression (8-step teacher march, every hop)287+.64*+1.04*+.73*+.0580.93.115
MeanFlow-style (two-time map, Eulerian identity)193+1.04*+2.00*+1.91*+.69*69.98.102

PTD is more dynamic than LMD, sharper and far more diverse; LMD is more natural by .11. LMD's clips carry 37% of the teacher's motion energy (PTD 118%), and its three seeds of a prompt are close variants of one shot (seed diversity .074, the DMD student's level) (videos). PTD keeps the motion and the diversity at a third of LMD's cost per update and with an unchanged deployed network. Regressing teacher-integrated endpoints (Euler row) loses on every quality axis, and the MeanFlow-style two-time map does not converge in 64 updates.

π-Flow. π-Flow deploys a distilled policy that is integrated with sub-steps at inference (App. A.4). LMD shares PTD's training objective, so it is the matched baseline; App. A.4 now states π-Flow's position explicitly.

W2: end-to-end K=0. The learned curve helps on both backbones we tested.

MiniMax-H3, end to end. K=0 trains with the paper's H3 recipe (150 updates) without the curvature head, so the teacher queries lie on the chord; the comparison is the paper model (K=8):

paper model − K=0 (* significant)dynamicstaticnaturalnessdiversity
all 100 prompts+.09*+.03+.03−.07
50 action prompts+.14*+.01−.01−.12*

K=0 also adds fine-scale flicker that the 12-frame judge cannot see: on the full videos its finest-band power is 120% of the teacher's, the paper model's 105% (videos).

Wan2.1-1.3B. K=1 (learned curve) against K=0 (straight chord) with identical prompts, seed, teacher, query budget and updates:

Wan2.1-1.3B (* significant)dynamicstaticnaturalnessdiversityVBenchdyn. degree
K=1 − K=0 (Astra Δ)+.05+.10*+.21*.0083.84 vs. 83.7176.7 vs. 74.7

The curve makes the student more natural and sharper; K=0 shows red and blue color artifacts that the curve does not (videos). The revision reports the LMD and MeanFlow baselines in Sec. 4.4 (Table 2) and App. A.4, and both K=0 comparisons in Sec. 4.4 and App. D.5 (Table 15).

Comments on the paper

3 comments
AuthorsOpenReview ↗

General response (1/2): the K=0 baseline and the established alternatives

We thank the reviewers. The reviews converge on five concerns. We answer each with new controlled experiments run with the paper's recipes and code, and we link side-by-side videos for every comparison. Protocol throughout: the paper's judging protocol (Sec. 4.1, App. C: four Astra axes on 100 prompts × 3 seeds); Δ = PTD − other, so a positive value favors PTD. Part 2/2 covers the remaining concerns, a label correction and the revision.

1. Does the learned curvature improve video quality, or is supervising the student along its own path enough? An end-to-end K=0 baseline is missing.

The curve helps on both backbones. Trained identically but with the teacher queried on the straight chord (K=0), the student is less dynamic and flickers on MiniMax-H3, and is less natural and less sharp, with color artifacts, on Wan2.1-1.3B.

MiniMax-H3, end to end. K=0 trains with the paper's H3 recipe (150 updates) without the curvature head; the comparison is the paper model (K=8):

paper model − K=0 (* significant)dynamicstaticnaturalnessdiversity
all 100 prompts+.09*+.03+.03−.07
50 action prompts+.14*+.01−.01−.12*

K=0 also adds fine-scale flicker that the 12-frame judge cannot see: on the full videos its finest-band power is 120% of the teacher's, the paper model's 105% (videos).

Wan2.1-1.3B. K=1 (learned curve) against K=0 (straight chord) with identical prompts, seed, teacher, query budget and updates:

Wan2.1-1.3B (* significant)dynamicstaticnaturalnessdiversityVBenchdyn. degree
K=1 − K=0 (Astra Δ)+.05+.10*+.21*.0083.84 vs. 83.7176.7 vs. 74.7

K=0 shows red and blue color artifacts that the curve does not (videos).

2. How does PTD compare with established alternatives, LMD, MeanFlow-based distillation and multi-step Euler endpoint regression, at comparable training and inference budgets?

Trained with PTD's recipe on Wan2.1-14B, every alternative loses to PTD on every quality axis except LMD's naturalness, and PTD alone keeps the seed diversity. Each row uses the paper's Wan recipe (App. B.2: batch 256, 64 updates, same teacher, data and 4-NFE inference) and changes only the training target or the map (videos):

Δ = PTD@64 − alternative (* significant)teacher velocities / samples / updatedynamicstaticnatural.diversityVBenchdyn. degreeseed div. (V-JEPA 2)
PTD@64 (reference)211884.5776.1.122
Euler endpoint regression (8-step teacher march, every hop)8287+.64*+1.04*+.73*+.0580.9356.4.115
LMD (conventional two-time map, same 2 target times)2342+.32*+.08*−.11*+.86*83.1848.3.074
MeanFlow-style (two-time map, Eulerian identity)1193+1.04*+2.00*+1.91*+.69*69.9837.5.102
MeanFlow-style, MeanFlow's default sampling (75% equal-time samples)1192+.73*+1.43*+1.34*+.50*

Regressing teacher-integrated endpoints (Euler row) costs 2.4× more per update and loses the fine detail that Fig. 6 shows. LMD is the closest alternative: it is more natural by .11, at 2.9× the cost per update, and its clips carry 37% of the teacher's motion energy (PTD 118%), with three seeds of a prompt that are close variants of one shot (seed diversity .074, the DMD student's level). The MeanFlow-style map does not converge in 64 updates, with or without MeanFlow's default equal-time samples. What PTD changes relative to LMD is what is deployed: the target time enters only through a training-only head, so inference runs the unchanged backbone, and all target-time queries share one student pass without Jacobian-vector products (App. A.4).

AuthorsOpenReview ↗

General response (2/2): curvature vs. quality, the shared-LoRA ablation, DMD across backbones; a label correction; the revision

3. Does the learned curvature relate to endpoint error and final video quality? Figs. 5 and 8 measure only curvature alignment, and Fig. 6 is qualitative.

Across every saved checkpoint of the paper's PTD-14B run, the learned bend grows steadily and aligns with the teacher's, while the endpoint error stays flat and does not track video quality. The curve's effect on quality comes from where it places the teacher queries (concern 1).

updates326496128160192224
hop-0 bend ratio / teacher-aligned bend.05 / .010.08 / .014.11 / .021.14 / .033.18 / .040.19 / .046.20 / .054
endpoint error, hop 0 / hop 3.62 / .114.72 / .106.75 / .106.66 / .128.74 / .122.70 / .129.70 / .121
VBench84.4684.5783.2984.0082.9283.9183.86

The student that regresses teacher-integrated endpoints (Euler row in part 1/2) has the lowest endpoint error of all the variants we trained (hop 0: .46 vs. .72 for PTD@64) and a static quality 1.04 below PTD@64.

4. Do the gains on MiniMax-H3 come from PTD or from training a separate LoRA per step?

From PTD. With one shared rank-128 LoRA for all four steps, exactly LightX2V Turbo's adapter budget, PTD keeps its gains over Turbo and stays within .05 of the per-hop paper model on every quality axis.

MiniMax-H3 (* significant)adaptersdynamicstaticnaturalnessdiversity
Paper model − Turbo (re-judged with the new rows)4 × rank-128+.11−.02+.14*+1.30*
Shared-LoRA PTD − Turbo1 × rank-128+.04−.08*+.09*+1.50*
Shared-LoRA PTD − paper model1 vs. 4 × rank-128−.04−.05*−.03+.12*

5. Do the gains over DMD hold beyond one released H3 checkpoint, and do they come from the trajectory supervision rather than from the final-hop teacher integration?

They hold on a second backbone and they survive the ablations. Against LightX2V's released four-step Wan2.1-T2V-14B DMD student, PTD@64 repeats the H3 pattern; removing the final-hop march costs quality and leaves diversity unchanged; and the LMD control (concern 2) shows that the single-time deployed map, with the target time confined to the training-only head, is the part of the trajectory supervision that keeps the diversity.

Δ = PTD − other (* significant)dynamicstaticnaturalnessdiversityVBenchseed div. (V-JEPA 2)
Wan2.1-14B: PTD@64 − LightX2V 4-step DMD student+.11−.07+.40*+.83*84.57 vs. 83.97.122 vs. .078
MiniMax-H3: no final march (2 queries on all hops) − paper model−.56*−.50*−.83*+.06

VBench does not separate PTD from the DMD student on Wan (PTD 82.92 to 84.57 across checkpoints); seed diversity confirms the gap (videos).

We sincerely thank all reviewers for the they put into reading our paper and into the discussion. Their concrete suggestions made the paper clearer and its evidence stronger, and we are glad that our responses addressed their concerns and that every reviewer maintained or raised their score to a positive assessment. We have updated the PDF with the new experiments and the ablations.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.