Apple and Cornell FastFrames Achieves 36ms Video Generation Latency on Single GPU

·2026.11.21 04:31

Key point

The 1.3B parameter model runs at 28 requests per second on a single NVIDIA B200 GPU, matching the quality of state-of-the-art models while being 8 times faster.

Details

  • Apple and Cornell University researchers developed FastFrames, an autoregressive video diffusion model optimized for interactive latency.
  • FastFrames-S (1.3B) achieves 36ms per frame (28 RPS) on a single NVIDIA B200 GPU, an 8x speedup over the CF++ baseline (272ms) with equivalent visual quality.
  • FastFrames-L (14B) achieves 45ms per frame (22 RPS) on an 8x B200 GPU node, offering superior realism and prompt adherence compared to Krea-Realtime (1,675ms on 1 GPU).
  • Profiling revealed that VAE decoding and KV cache updates account for over 50% of inference latency in state-of-the-art models.
  • The system replaces these operations with lightweight alternatives and applies system-level optimizations to accelerate diffusion sampling.
  • The 14B model generates more realistic lighting, consistent composition, and legible text rendering compared to 1.3B models.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.