Aryan V S

Iterative sampling process, large denoising networks -> prohibitively high computational requirements for practical deployment

For $T$ sampling steps, $F$ frames, and per-frame feature length $L$, total complexity is $T \times F^2 \times L^2$.

Most existing models optimize the sampling steps but overlook other expensive factors. Authors note that optimization along the frame dimension, which contributes, $O(f^2)$ complexity, remains underexplored.

Recent approaches like CausVid employs asymmetric structural distillation strategy - bidirectional-attention teacher diffusion model and causal student model--mainly reduces cost by fewer sampling steps.

Authors find that attention to first frame and neighbouring frames is substantially higher than to distant ones across multiple foundation video-generation models--Wan, Hunyuan Video and CogVideoX.