Training-free method improves reuse of video diffusion adapters
DART: Distillation-Aware Reparameterization for Training-Free LoRA Reuse in Few-Step Video Diffusion Models
Artificial Intelligence
Summary
Making videos using AI can be slow and expensive, so reusing smaller parts called LoRAs speeds things up. But using a LoRA made for a longer process in a shorter one can cause problems or reduce quality. The paper introduces DART, a way to adjust LoRAs without extra training so they work better in shorter video generation steps. The method improves video quality and keeps the adapters’ functions more intact, helping avoid negative effects when reusing them.
What this means in practice
- •For video ai developers: Improve efficiency in video generation by reusing adapters trained on longer schedules for faster, fewer-step video diffusion.
- •For machine learning engineers: Calibrate low-rank adapters to maintain quality and function when transferring models trained on different video generation timelines.
Authors
Shihong Li, Juntao Xu, JinCao, Maowen Tang, Jun Huang, Jintao Li
Abstract
Step distillation reduces the cost of video generation, but reusing a LoRA trained for a longer trajectory can alter its functional effect or degrade target quality. Static parameter compatibility offers one perspective on this problem; our observations show that similar measured geometry can coexist with different adapter behavior under a shortened denoising schedule. We propose DART, a training-free method that combines low-rank coordinate transport with target-schedule response calibration using forward evaluations and no source training videos. On a four-step Wan2.2 target, DART-F improves the joint quality score from 0.9029 to 0.9227 and changes macro functional retention from -0.4644 to +0.1349. Component analysis shows that calibration accounts for most of the quality improvement, while coordinate transport provides complementary gains when combined with calibration. Adapter-level results reveal positive functional effects for some adapters and strong attenuation with reduced negative functional effects for others. Evaluations on two additional targets show the same aggregate trend. These results motivate evaluating distilled-model LoRA reuse jointly through functional preservation and negative-transfer avoidance, without assuming recovery for every adapter.