Source-linked AI summary
Anchor Forcing: Anchor Memory and Tri-Region RoPE for Interactive Streaming Video Diffusion
Yang Yang, Tianyi Zhang, Wei Huang, Jinwei Chen, Boxi Wu, Xiaofei He, Deng Cai, Bo Li, Peng-Tao Jiang
TL;DR
Interactive long-video generation must handle prompt switches while preserving visual quality and coherent motion, but existing streaming methods lose semantic or recent cache cues and exceed pretrained positional regimes. Anchor Forcing combines anchor-guided re-cache with tri-region RoPE to address these issues, improving perceptual quality and motion dynamics over prior streaming baselines in interactive settings.
Problem
Interactive streaming methods struggle to preserve semantic context and recent latent cues at prompt switches, while unbounded indexing deviates from the pretrained bounded RoPE regime.
Method
Anchor Forcing warm-starts re-cache from anchor memory and uses region-specific, range-bounded tri-region RoPE to retain motion priors.
Results
Anchor Forcing improves perceptual quality and motion dynamics and consistently outperforms prior streaming baselines in interactive settings.
Takeaways & Limitations
The framework achieves state-of-the-art results for interactive long-video generation on VBench benchmarks.
Abstract
from arXiv · showhide
Interactive long video generation requires prompt switching to introduce new subjects or events, while maintaining perceptual fidelity and coherent motion over extended horizons. Recent distilled streaming video diffusion models reuse a rolling KV cache for long-range generation, enabling prompt-switch interaction through re-cache at each switch. However, existing streaming methods still exhibit progressive quality degradation and weakened motion dynamics. We identify two failure modes specific to interactive streaming generation: (i) at each prompt switch, current cache maintenance cannot simultaneously retain KV-based semantic context and recent latent cues, resulting in weak boundary conditioning and reduced perceptual quality; and (ii) during distillation, unbounded time indexing induces a positional distribution shift from the pretrained backbone's bounded RoPE regime, weakening pretrained motion priors and long-horizon motion retention. To address these issues, we propose \textbf{Anchor Forcing}, a cache-centric framework with two designs. First, an anchor-guided re-cache mechanism stores KV states in anchor caches and warm-starts re-cache from these anchors at each prompt switch, reducing post-switch evidence loss and stabilizing perceptual quality. Second, a tri-region RoPE with region-specific reference origins, together with RoPE re-alignment distillation, reconciles unbounded streaming indices with the pretrained RoPE regime to better retain motion priors. Experiments on long videos show that our method improves perceptual quality and motion metrics over prior streaming baselines in interactive settings. Project page: https://github.com/vivoCameraResearch/Anchor-Forcing
1 Introduction
Interactive streaming video generation must support frequent prompt switches while preserving long-horizon visual quality and motion. Anchor Forcing addresses cache evidence loss and positional mismatch with anchor-guided re-cache and tri-region RoPE, improving interactive results.
- Motivation: Long-horizon interactive generation requires efficient streaming because dense spatiotemporal attention becomes computationally prohibitive.Users also expect low latency and frequent edits.
- Existing Streaming Models: Rolling KV caches enable efficient long-range context reuse and prompt-switch interaction in distilled streaming video diffusion.These systems combine short-window causal attention with sink tokens and cache updates.
- Failure Modes: Re-cache methods discard earlier semantic KV states at prompt switches, weakening cross-segment continuity and post-switch perceptual quality.This leaves less semantic context available for conditioning after a prompt change.
- Anchor Forcing: Anchor-guided re-cache builds anchor memory from junction, sink, and local caches to warm-start subsequent generation and reduce post-switch evidence loss.The mechanism is designed to stabilize quality after prompt changes.
- Anchor Forcing: Tri-region RoPE assigns bounded, region-specific positional encodings to preserve pretrained positional behavior and improve long-horizon motion-prior retention.The three regions partition anchor memory and support adaptive cross-region retrieval during distillation.
- Results: Anchor Forcing improves perceptual quality and motion dynamics on VBench and achieves state-of-the-art interactive results over prior streaming baselines.The introduction consistently reports these gains across the method summary and contribution statement.
2 Related Work
Long-video research has pursued extended context, spectral and state-space memory, and causal streaming with KV-cache reuse. Despite efficiency and temporal-length gains, interactive systems still struggle with prompt-switch cache maintenance and bounded positional modeling.
- Long-Context Generation: Long-video methods extend temporal context through larger spatiotemporal windows, error recycling, sparse routing, spectral fusion, random-mask conditioning, and state-space architectures.These approaches target longer generation or more efficient long-range modeling.
- Streaming Generation: Streaming autoregressive methods use causal temporal modeling and temporal KV-cache reuse to improve efficiency and support long-horizon generation.Representative work includes CausVid, Self-Forcing, and Infinity-RoPE.
- Remaining Challenges: Interactive streaming still fails to jointly retain semantic KV context and recent latent cues at prompt switches, weakening boundary conditioning and perceptual quality.Unbounded distillation-time indexing also departs from the pretrained bounded RoPE regime and undermines motion-prior transfer.
3 Method
Anchor Forcing combines anchor-guided KV re-caching with tri-region RoPE to preserve semantic evidence at prompt switches and keep streaming positions within the pretrained range. The method organizes boundary, local, and sink context while using bounded relative positions to support long-horizon motion retention.
- Streaming autoregressive video diffusion: Streaming autoregressive diffusion generates temporal chunks with causal attention while reusing local and sink KV caches for efficient long-range conditioning.The model conditions on recent latent tokens, persistent sink context, and the current text prompt.
- Problem formulation: At prompt switches, conventional re-cache evicts prior KV states, removing informative preceding context and weakening cross-switch semantic continuity.This loss of semantic evidence reduces effective conditioning around interaction boundaries and degrades perceptual quality.
- Anchor-guided re-cache: Anchor-guided re-cache stores boundary-adjacent junction KV states and combines them with sink and rolling local caches to preserve newly introduced interaction cues.Junction caches are refreshed after each boundary and remain fixed until the next boundary, while local caches retain recent short-range context.
- Anchor-guided re-cache: The anchor memory activates junction evidence after those frames leave the local window, keeping boundary information persistently accessible during later segments.The activation condition prevents redundant use while the junction frames remain in the rolling local context.
- Tri-region RoPE: Unbounded global RoPE indices exceed the pretrained positional range, causing distribution shift that weakens pretrained motion priors and long-horizon motion retention.The generator is initialized from CausVid, whose ODE trajectories are capped at 21 latent frames.
- Tri-region RoPE: Tri-region RoPE assigns bounded, region-specific relative indices to sink, local, and junction caches, keeping attention positions within Pmax while enabling complementary cross-region retrieval.The local region uses a rolling relative indexing scheme, and junction indices are placed immediately before the local cache; coupled distillation improves motion dynamics.
4 Experiments
Experiments evaluate Anchor Forcing in interactive and non-interactive streaming settings, including long-horizon comparisons and component ablations. The results show stronger prompt adherence, perceptual quality, and motion dynamics, with tri-region RoPE improving motion and anchor-guided re-cache stabilizing boundary conditioning.
- Interactive Generation: Anchor Forcing achieves the best dynamic degree and quality score among methods in 60-second interactive generation, while maintaining stronger semantic adherence in later segments.Its qualitative results also show more accurate prompt-induced interactions, smoother transitions, and higher visual fidelity across switches.
- Interactive Generation: Anchor Forcing follows newly introduced subjects and actions more reliably than baselines, which often produce weak or ambiguous prompt-aligned interactions.The comparison highlights clearer realization of the “young couple” and later sushi-related actions.
- Non-Interactive Generation: Anchor Forcing ranks first overall for 30-second single-prompt generation and achieves the best total and quality scores on VBenchLong.It also reports the highest imaging quality and temporal style, indicating stronger long-horizon visual and stylistic stability.
- Ablation Studies: Tri-region RoPE increases dynamic degree from 26.37 to 74.83 by using regionwise, range-bounded indexing within the pretrained RoPE regime.The component improves motion coherence but can still leave appearance degradation and inconsistent prompt responsiveness, motivating anchor-guided re-cache.
- Ablation Studies: Adding anchor-guided re-cache raises quality from 81.72 to 82.25 and CLIP from 23.71 to 23.85 on top of tri-region RoPE.The mechanism improves prompt-faithful content updates across segments while preserving strong dynamics.
- Ablation Studies: Compared with baseline re-cache and Flush, anchor-guided re-cache achieves the best reported CLIP, dynamic degree, and quality results.The broader ablations characterize tri-region RoPE as critical for motion-prior retention and anchor-guided re-cache as complementary for boundary conditioning and content fidelity.
5 Conclusion
The paper presents Anchor Forcing as a cache-centric framework for interactive long-video generation under streaming constraints. It combines anchor-guided re-cache with tri-region RoPE to improve perceptual quality and motion dynamics over prior streaming baselines.
- Conclusion: Anchor Forcing combines anchor-guided re-cache for post-switch conditioning with tri-region RoPE for retaining pretrained motion priors over long horizons.The framework achieves state-of-the-art results on VBench in interactive settings.