We introduce DART, a training‑free reparameterization that preserves LoRA functionality when the denoising schedule is shortened. Step distillation reduces video generation cost by cutting diffusion steps, but directly reusing a LoRA trained on a long trajectory often leads to functional drift or quality loss. Our observations reveal that even when measured geometry is similar, adapters can behave differently under a reduced schedule, indicating that static parameter compatibility alone is insufficient.
DART consists of two components: (1) low‑rank coordinate transport, which maps source LoRA parameters into the coordinate space of the target schedule; (2) target‑schedule response calibration, which evaluates the forward output on the target step count and corrects it without any source videos or extra training. Calibration uses the forward error $\Delta = f_{\text{target}}(x) - f_{\text{source}}(x)$ for a linear adjustment.
On a four‑step Wan2.2 target, DART‑F raises the joint quality score from $0.9029$ to $0.9227$ and shifts macro functional retention from $-0.4644$ to $+0.1349$. Component analysis shows calibration accounts for most of the gain, while coordinate transport adds complementary improvements when combined with calibration. Adapter‑level results indicate positive functional effects for some adapters and a strong attenuation of negative effects for others. Evaluations on two additional targets exhibit the same overall upward trend.
These findings suggest that assessing distilled‑model LoRA reuse should jointly consider functional preservation and negative‑transfer avoidance, rather than assuming every adapter can fully recover its original performance.
Review