NVIDIA · 2026-05-13 · notable
NVIDIA AnyFlow — Any-Step Video Diffusion Distillation With Flow-Map Transition Learning Beats Consistency Methods at 1.3B and 14B Scale
AnyFlow is NVIDIA's distillation framework for video diffusion. Instead of locking a student model to a fixed step count, it learns flow-map transitions over arbitrary time intervals so quality scales with whatever step budget you spend at inference.
Train one video diffusion student that runs well at 1, 4, or 32 steps — no separate model per step budget.
Key specs
| Model sizes | 1.3B and 14B params |
|---|---|
| Model variants | 4 |
| Hf day upvotes | 75 |
| Base backbones | Wan2.1 (bidirectional) and FAR (causal) |
What is it?
AnyFlow is a distillation recipe for text-to-video diffusion. The team distills four open Wan2.1 variants (1.3B and 14B, in both bidirectional and FAR-style causal layouts) into students that keep their quality across arbitrary sampling step counts, not just a fixed few.
How does it work?
Consistency distillation collapses the original probability-flow ODE trajectory into a single endpoint map (z_t → z_0), which hurts quality when you actually sample with more steps. AnyFlow shifts the distillation target to flow-map transition learning (z_t → z_r) over arbitrary intervals, and pairs it with a backward-simulation trick so the student stays on-policy during training. The result is a student whose performance keeps improving as you give it more compute at inference.
Why does it matter?
Few-step video diffusion has been a brittle trade-off: faster sampling came at the cost of giving up the test-time scaling curve that ODE samplers enjoyed. AnyFlow gives back that scaling axis while staying competitive in the 1-to-4-step regime, and ships open-weights checkpoints under Apache 2.0 instead of a paper-only release.
Who is it for?
Video model researchers, anyone serving Wan2.1-class T2V models, and teams that want to dial inference cost up and down at runtime.
Try it
huggingface.co/collections/nvidia/anyflow