Diffusion-based video generation is computationally heavy because every denoising step processes a large number of tokens. Progressive resolution sampling, which lowers latent resolution early in the process, has helped for images but does not translate directly to video: the extra temporal dimension creates varying spatio-temporal demands across clips, and compressing only one dimension either limits speedup or hurts quality.
DVG, proposed by Shikang Zheng and colleagues, treats this as a joint allocation problem. Instead of fixing a compression schedule, it dynamically distributes computation across both time and space, choosing content-aware acceleration strategies per video. The framework requires no manual tuning or retraining, so it can be dropped into existing models.
On HunyuanVideo and HunyuanVideo-1.5, the authors report up to 7x speedup with near-lossless quality, and up to 18x when combined with distillation. The code is publicly available, which should make it easy for others to test the method on their own systems.