Accelerating Video Diffusion via Training-Free Trajectory Routing
TRACK routes easy denoising steps to a smaller model and reports up to 2.73x faster video diffusion without retraining.
The method calibrates where a small model disagrees least with a large reference model, then uses that map to decide which model runs at each diffusion step. Quality-sensitive steps stay on the large model; lower-disagreement steps switch to the small one. The paper says inference runs only one selected model per step, with no architecture, scheduler, or online dual-model changes. Reported speedups span Wan 2.1, Cosmos 3, TurboDiffusion, and FastVideo, with comparable aggregate quality and high diversity retention. ArXiv · AI/CL/LG's note
The method calibrates where a small model disagrees least with a large reference model, then uses that map to decide which model runs at each diffusion step. Quality-sensitive steps stay on the large model; lower-disagreement steps switch to the small one. The paper says inference runs only one selected model per step, with no architecture, scheduler, or online dual-model changes. Reported speedups span Wan 2.1, Cosmos 3, TurboDiffusion, and FastVideo, with comparable aggregate quality and high diversity retention. ArXiv · AI/CL/LG's note
score 5