In-Flight KV Cache with Clean Anchors for Faster Autoregressive Video Diffusion
FlashForward reuses denoising-stage KV caches already computed in flight, then steadies the result with sparse clean anchor latents.
The paper targets few-step autoregressive video diffusion, where prior methods spend extra forwards rebuilding cleaner KV cache history. FlashForward avoids those cache-update-only passes and lets chunks move through different denoising stages concurrently across GPUs. The authors say the noisy stage-matched history can cause appearance and motion drift, so they add clean anchors for longer-range structural guidance. Reported speedups reach 1.16-1.69x over HiAR and 1.42-2.92x over Self-Forcing on 20-second-plus 16 FPS videos, with VBench gains for the 1.3B 480p setting. HF Daily Papers' note
The paper targets few-step autoregressive video diffusion, where prior methods spend extra forwards rebuilding cleaner KV cache history. FlashForward avoids those cache-update-only passes and lets chunks move through different denoising stages concurrently across GPUs. The authors say the noisy stage-matched history can cause appearance and motion drift, so they add clean anchors for longer-range structural guidance. Reported speedups reach 1.16-1.69x over HiAR and 1.42-2.92x over Self-Forcing on 20-second-plus 16 FPS videos, with VBench gains for the 1.3B 480p setting. HF Daily Papers' note
score 4