Interactive video generation is often bottlenecked by more than the diffusion model itself: decoding pixels and updating the KV cache can consume over half of end-to-end latency. This model’s design treats those stages as optimization targets too, helping a 14B generator reach interactive rates—provided the workload has substantial GPU capacity.
Key Capabilities
- Two-step autoregressive sampling generates one latent frame per step, enabling ongoing video generation rather than waiting for a complete clip.
- The 14B variant reaches about 45 ms per frame, or 22 responses per second, on a node with eight B200 GPUs. That puts it below the 50 ms interactive-latency target in the reported setup.
- Its Wan2.1-14B foundation is distilled into a causal model, with a lightweight pixel decoder and KV-refinement network trained around the generator. The released training checkpoints make the distillation stages available for inspection.
- It supports text-to-video and is also part of work on camera-conditioned interactive generation.
Who It’s For
A good fit for researchers and teams exploring low-latency video generation who can run multi-GPU experiments and care about frame-by-frame interaction. Look elsewhere if you need a practical single-GPU deployment: the reported FastFrames-L latency rises to about 158 ms per frame on one B200, and its resource requirements are significant.