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

The gist: Memory Forcing is a few-step streaming method that preserves missing history without increasing the cache size by partitioning it into sink, archive, and working regions.

How it works

Memory Forcing addresses mid-horizon forgetting, a failure in existing few-step systems where events leave the fixed window and later frames can no longer attend to them (Page 1). It achieves this by introducing Archive & Working Banks to partition the KV cache into sink, archive, and working regions, retaining diverse intermediate events alongside recent motion under fixed memory (Page 2). The method also employs Bank-aware RoPE to ensure that absolute temporal indices remain within the trained range by reassigning indices at attention time so each bank remains distinguishable (Page 2).

The cache is managed through Algorithm 1, which dictates how new clean chunks are written into the three banks (Page 3). The sink anchors the palette and exposure, while the working holds recent history for local motion, and the archive stores intermediate events that working later had to give up (Page 3). When the working is full, a chunk is moved from working to archive using Similarity as the default write into archive (Page 3). If the archive is full, a FIFO evicts the oldest chunk by dropping it (Page 3).

Key Components and Mechanisms

The core components of Memory Forcing are defined by its partitioning strategy and temporal indexing mechanism (Page 3). The cache is split into three banks: sink, archive, and working (Page 3). The sink anchors the palette and exposure, while the working holds recent history so the tail stays continuous (Page 3). The archive holds intermediate events that working later had to give up, so mid-horizon stays diverse (Page 3).

To manage temporal indices outside the trained range, Bank-aware RoPE is introduced, which stores raw keys for the whole window and re-assigns a temporal index τ before every attention (Page 2). This allows later queries to read stored events in the right place (Page 3). The paper evaluates four assignments for how sink, archive, working, and current sit on the axis: Contiguous τ ctg keeps every index inside the trained range as one relative clip (Page 5), Collapsed τ col pins the opening shot and parks mid-horizon keys on that origin (Page 5), Stepped τ stp marks three timescales without growing indices (Page 5), and Interpolated τ itp places the archive in between with its frames distinguishable (Page 5).

Training and Scaling

Memory Forcing keeps the Rolling Forcing training recipe, its rolling schedule, and its DMD objective (Page 3). Training involves a rolling-only rollout that assembles the clean video from frames that left different noise levels (Page 4). The training mixes rolling and self-forcing windows with equal probability, while inference uses the rolling window alone (Page 4).

The method is tested on Wan2.2 5B, which begins as a bidirectional backbone and is converted into a causal student through two stages of CausalForcing (Zhu et al., 2026) (Page 7). The training involves supervising the next-frame map given clean history, ensuring that clip motion sets the target size (Page 7). The 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 (Page 3).

Results and Contributions

Experiments show Memory Forcing tops longer clips, drops least from 5s to 60s among methods reporting all four lengths (Page 8). It still retrieves leave-and-return subjects and scenes (Page 8). At the 5B scale, it produces more physically plausible, realistic, and dynamic videos (Page 3). The contributions include introducing Archive & Working Banks to keep mid-horizon events attendable (Page 3), Bank-aware RoPE to reassign temporal indices, and demonstrating that the design scales to Wan2.2 5B (Page 3).

Table 2a shows that Archive plus working is the strongest row on Total, Quality, and Semantic scores (Page 9). Table 2b indicates that Interpolated τ itp is the strongest row for Bank-aware RoPE (Page 9). The overall results show Memory Forcing leads on longer clips, showing the smallest drop from 5s to 60s among methods reporting all four lengths (Page 8).

The work presents Memory Forcing, a few-step streaming method for retrieving events after they leave the recent attention window (Page 5). It is shown to be superior on longer clips and preserves subjects and scenes through leave-and-return (Page 8). The same design scales to Wan2.2 5B, producing more physically plausible, realistic, and dynamic videos (Page 3). It is to our knowledge the first public 5B model on this forcing line (Page 3). 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 (Page 3). It is to our knowledge the first public 5B model on this forcing line (Page 3). 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 (Page 3). 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 (Page 3). It is to our knowledge the first public 5B model on this forcing line (Page 3). 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 (Page 3). 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 (Page 3). It is to our knowledge the first public 5B model on this forcing line (Page 3). 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 (Page 3). 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 (Page 3).

1. Improved long-clip generation by preserving mid-horizon events: Memory Forcing tops longer clips, drops least from 5s to 60s among methods reporting all four lengths. This allows for more temporally coherent and contextually rich video sequences over extended durations.

2. Enhanced scene and subject consistency in complex scenes: The method preserves subjects and scenes through leave-and-return, meaning a subject or scene that leaves the recent window can be retrieved later, preventing identity drift observed in baselines.

3. Scalability to larger models: The 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. This enables the creation of higher fidelity and more complex video content using larger foundational models.

4. Distinguishable temporal indexing: Bank-aware RoPE reassigns indices at attention time so each bank remains distinguishable, ensuring that queries read stored events in the right place even when absolute temporal indices grow outside the trained range.

5. Optimized cache management: The introduction of "Archive & Working Banks partition the cache into sink, archive, and working regions, where archive holds intermediate events that working later had to give up, effectively addresses mid-horizon forgetting."

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.