Parametric trajectory distillation improves fast video generation quality
Parametric Trajectory Distillation for Few-Step Video Generation
Computer Vision and Pattern Recognition
Summary
Generating videos step-by-step with current AI methods takes a lot of time and computing power. The authors introduce a way called Parametric Trajectory Distillation that helps these models learn to create videos quickly in just a few steps without losing quality. They do this by letting the model follow smooth paths (polynomials) learned from a slower, more detailed version. This technique keeps motion and details clear while making the process faster. Tests show it produces more natural and diverse videos compared to existing fast-generation methods.
What this means in practice
- •For video content creators: Produce high-quality short videos quickly with fewer computation steps using the new parametric trajectory approach.
- •For mobile app developers: Integrate faster video generation models that maintain natural motion for apps requiring on-device or low-latency video synthesis.
Authors
Lan Feng, Peter Karkus, Maximilian Igl, Julius Berner, Yuxiao Chen, Shuhan Tan, Alexandre Alahi, Boris Ivanovic, Marco Pavone
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. Project page: https://alan-lanfeng.github.io/PTD/.