Memory Forcing: Attendable Mid-Horizon History for Streaming Video Generation

2026-10-08

AI Summary

This study addresses the degradation in long-segment quality and subject inconsistency in streaming video generation, which arises from fixed KV cache limits causing models to forget mid-term history. To overcome this, we propose Memory Forcing, a framework that introduces a novel partitioned caching mechanism comprising Archive and Working Banks to retain mid-term events. Furthermore, we design Bank-aware RoPE, a dynamic positional encoding reallocation strategy that preserves attention access to mid-term history without increasing memory overhead. Evaluated on a 1.3B autoregressive video diffusion model, our approach significantly mitigates long-horizon performance degradation and enhances subject consistency. Scaling the framework to the Wan2.2 5B model enables the generation of long-duration dynamic videos with high physical realism.

Abstract

Autoregressive video diffusion enables causal video streaming without a bidirectional pass over the full clip, but existing few-step systems usually retain only the opening and most recent frames in a fixed-size KV cache. Once an event leaves this window, later frames can no longer attend to it, a failure we term mid-horizon forgetting. We present Memory Forcing, a few-step streaming method that preserves this missing history without increasing the cache size. Its Archive \& Working Banks partition the cache into sink, archive, and working regions, retaining diverse intermediate events alongside recent motion under fixed memory. Because absolute temporal indices drift outside the training range, Bank-aware RoPE reassigns indices at attention time so each bank remains distinguishable. At 1.3B, Memory Forcing leads on longer clips, shows the smallest drop from 5s to 60s among methods reporting all four lengths, and preserves subjects and scenes through leave-and-return. The same design scales to Wan2.2 5B, producing more physically plausible, realistic, and dynamic videos and, to our knowledge, the first public 5B model on this forcing line.

Problem

• streaming video generation
• autoregressive video diffusion
• mid-horizon forgetting
• KV cache
• temporal consistency

Innovation

• Memory Forcing
• Streaming Video Generation
• Autoregressive Video Diffusion
• Bank-aware RoPE
• Mid-Horizon Forgetting