Decoupled Long Video Generation with Chain of Training-free Diffusion Experts

Video generation models hold substantial potential in areas such as filmmaking. However, current video diffusion models struggle with poor quality and slow speed due to extreme complexity of video generation task. In this paper, we propose ConFiner, an efficient video generation framework that decouples video generation into three easier subtasks: structure control, spatial refinement and temporal refinement. It can generate high-quality videos with chain of off-the-shelf diffusion model experts, each expert responsible for a decoupled subtask. During the refinement, we introduce coordinated denoising, which can merge multiple diffusion experts' capabilities into a single sampling. Furthermore, we design ConFiner-Long framework, which can generate long coherent video with three constraint strategies on ConFiner. Experimental results indicate that with only 20% of the inference cost, our ConFiner achieves higher average score than representative models like Wan2.1 and HunyuanVideo on Vbench. And ConFiner-Long can generate high-quality and coherent videos with infinite length.