Few-step distillation is a capacity allocation problem
Video diffusion and flow models generate by evaluating the network many times in sequence. Few-step distillation compresses that process into a handful of student evaluations, often keeping the teacher's backbone, so the student can have the same parameter count but far less sequential computation with which to reproduce a much longer generation process. The paper frames this as capacity allocation: supervision must help the student spend its limited computation on the parts of the teacher's path it can actually represent.
The failure mode is specific. Existing trajectory objectives train the student to reproduce the teacher's full noise-to-sample trajectory, which preserves motion and diversity but is mode-covering: when a few-step student cannot represent complex teacher transitions, regression pulls predictions toward an average of the targets and blurs fine detail. The paper notes that teacher transitions are highly curved at high noise, exactly where this averaging pressure is strongest.
Teacher queries follow the student's own curve
PTD changes what the student is asked to output. A single student evaluation predicts a polynomial approximation of a teacher trajectory segment, and teacher velocities are then queried along this predicted path rather than along the teacher's original path. The paper describes the design as analogous to distribution matching, which evaluates the teacher only on the student's own samples, while retaining the noise-to-sample correspondence that trajectory supervision provides.
The curve has two outputs trained by one velocity-regression loss. Its mean velocity directly supplies the endpoint update, and a curvature head supplies the polynomial's shape. Both are trained without Jacobian-vector products and without an auxiliary critic, which the authors present as a simple recipe for scaling trajectory distillation to large backbones, including parameter-efficient LoRA fine-tuning.
The curvature head is a training scaffold only
The curvature head exists to shape the supervision used to learn the backbone's velocity prediction; it is discarded at inference. Sampling follows x_{k+1} = x_k + Δ_k h_θ(x_k, t_k | y) for N student evaluations, with no teacher queries and no interior-curve evaluation. Each deployed hop retains the original backbone topology and output dimensions.
On MiniMax-H3, the per-hop LoRA adapters, including the adapted velocity projection, can be merged into the original projection shape, so the deployed model keeps its original layer types and output dimensions. On Wan2.1-14B, the fully fine-tuned backbone and dense mean-velocity projection are retained directly. In both cases the deployed model takes four steps and does not evaluate curvature coefficients.
Four-step PTD beats PDD on Wan2.1-14B
On Wan2.1-14B, four-step PTD is compared with a reproduction of Parallel Decoding Distillation (PDD), described as the best-performing trajectory-only method on this model, under the same training setting. PTD reaches an overall VBench score of 84.57 at iteration 64, while the PDD reproduction reaches 83.29, 83.67, and 82.83 at iterations 64, 96, and 256; PDD's best checkpoint, iteration 96, serves as the baseline for the Astra, human-preference, and VBench comparisons. The paper reports that PTD significantly improves dynamic quality and naturalness over PDD.
The authors also note that their PDD reproduction scores below the VBench values reported by the original PDD authors, which they attribute to differences in training prompt data and evaluation prompts and to VBench evaluation noise. That caveat qualifies the size of the margin, not the direction of the comparison within this training setting.
LoRA distillation of a 33B audio-video model
On the 33B audio-video MiniMax-H3, PTD is trained with per-hop LoRA and trajectory supervision alone into a four-step generator. Against LightX2V Turbo, described as a state-of-the-art four-step DMD method on the H3 Acceleration Arena, PTD significantly improves diversity (2.81 vs. 1.55) and naturalness as measured by GPT-6 Astra, while matching its dynamic and static quality. The reported checkpoint is iteration 150, with 124 frames and stereo 32-kHz audio.
The evaluation uses GPT-6 Astra as an automated judge over four dimensions: dynamic quality, static quality, naturalness, and diversity across three seeds. The paper states that these judgments are not human ratings, and audio is not evaluated. Diversity here is a judged difference across the three seeds for the same prompt, not a distribution-level metric.
Blinded votes favor PTD over both baselines
Blinded human votes give PTD 55.1% and 63.4% preference shares against PDD and LightX2V Turbo respectively. These are preference shares in a blinded comparison, not scores on the four Astra dimensions, so they should be read as a separate line of evidence about overall preference rather than as a substitute for the diversity or naturalness numbers.
The two comparisons also differ in setup: the PDD comparison is on Wan2.1-14B with full fine-tuning, while the LightX2V Turbo comparison is on MiniMax-H3 with LoRA training. The paper's claim is therefore a pair of backbone-specific wins, not a single ranking across models.
PTD learns a shrunken, teacher-aligned bend
A trajectory-shape probe measures how the student's predicted bend compares with the teacher's in the middle half of each hop. On Wan2.1-14B, the teacher bends most in the high-noise first hop: its velocity direction turns by 43.5° within hop 0, compared with 4.7° within the final hop. In that strongly curved hop, PTD has a larger bend magnitude ratio, direction cosine, and teacher-aligned projection than PDD at every shared checkpoint from 32 to 224 updates. At 64 updates, the magnitude ratio is 0.081 for PTD and 0.024 for PDD, with cosines of 0.171 and 0.139 and projections of 0.014 and 0.003.
The authors interpret this as the student learning a shrunken, teacher-aligned curve that keeps its targets learnable, even though the curve family can fit teacher trajectories almost exactly. On the smoother final hop, the teacher-aligned projections of the two methods differ by at most 0.005 at every shared checkpoint, which is consistent with the final hop being where the H3 recipe applies teacher marching.
Supervision should match what each hop can predict
The paper's diagnostics show that predictability varies sharply along the trajectory. Offline fits with K=2 Legendre terms explain 90% of final-hop velocity variation versus 46% on the hop nearest noise. Frozen-feature probes on H3 give held-out scores near 45% for ridge and 49–50% for Adam on the final hop, but only 7–11%, 14–15%, and approximately 3.5% on hops 0, 1, and 2 respectively.
This motivates the model-specific recipes. The H3 checkpoint adds teacher marching only on the final hop, where the endpoint is smoother, and reports that this improves visual detail at approximately 20% higher training compute than the all-two-query recipe. The Wan checkpoint uses the student-curve objective alone on all four hops. The authors report that marching every hop softens local detail on Wan2.1-1.3B, because a marched endpoint folds high-noise curvature into a target too hard for one student evaluation.
The practical priority that follows is not a universal recipe but a matching rule: query the teacher along the student's predicted curve where curvature is hard, and consider a marched target only where the hop is smooth enough for the student to learn it in one evaluation. The evidence for that split comes from Wan2.1-14B, MiniMax-H3, and a Wan2.1-1.3B ablation, so the specific hop assignment should be re-checked before transferring it to a different backbone or time grid.
Sources used in this answer
Parametric Trajectory Distillation for Few-Step Video Generation
PTD lets a few-step student predict polynomial teacher trajectory segments and learn from teacher guidance along its own path, beating PDD on Wan2.1-14B and LightX2V Turbo on MiniMax-H3 in the reported comparisons.