NVlabs/AnyFlow

What it solves

Traditional distilled video diffusion models are often locked into a fixed number of sampling steps, meaning they cannot adapt to different computational budgets. AnyFlow solves this by allowing a single model to generate high-quality video using an arbitrary number of steps, where quality improves stably as more steps are added.

How it works

AnyFlow is a video diffusion framework based on flow maps. It utilizes on-policy flow map distillation to enable any-step generation. The framework is compatible with both causal and bidirectional video diffusion architectures and can be scaled from 1.3B to 14B parameters.

Who it’s for

It is designed for researchers and developers working with video generation models who need flexibility in inference speed versus quality trade-offs.

Highlights

• Any-Step Generation: Adapts to any inference budget without being tied to a fixed step count.
• Versatile Tasks: Supports Text-to-Video, Image-to-Video, and Video-to-Video generation within a single causal model.
• Architecture Agnostic: Works with both causal and bidirectional video diffusion models.
• Scalable: Validated across model sizes ranging from 1.3B to 14B parameters.