Nvidia's AnyFlow adjusts diffusion steps dynamically for any compute budget
New any-step video framework from Nvidia runs on Wan2.1 at 1.3B or 14B parameters.
Deep Dive
AnyFlow is introduced as the first any-step video diffusion framework built on flow maps. It offers Wan2.1 model versions at 1.3B and 14B for both text-to-video and text/image-to-video, available via HuggingFace Diffusers. The original post notes uncertainty about ComfyUI support.
Key Points
- AnyFlow dynamically adjusts time steps during video diffusion, trading quality for speed based on compute budget.
- Supports Nvidia Wan2.1 models at 1.3B and 14B parameters for text-to-video and text/image-to-video.
- Available as diffusers on HuggingFace; not yet integrated into ComfyUI without manual modifications.
Why It Matters
Gives developers precise control over video generation speed and quality, enabling efficient deployment on varied hardware.