TetherCache: Stabilizing Autoregressive Long-Form Video Generation with Gated Recall and Trusted Alignment

2026-06-11 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

TetherCache is a training-free, plug-and-play cache management strategy designed to stabilize autoregressive long-form video generation. It addresses challenges in minute-level video generation, such as visual artifacts, quality degradation, and temporal drift, which arise from limited KV-cache budgets and context distribution shifts. TetherCache employs two mechanisms: GRAB (Gated Recall with Attention-Diversity Balancing), which selects diverse, informative long-range memory frames, and TAME (Trusted Alignment via Memory Editing), which aligns newly recalled memory tokens to a trusted context distribution to reduce feature pollution. Built on Self-Forcing, TetherCache consistently improves long-video generation quality on VBench-Long across 30s, 60s, and 240s settings, notably reducing quality drift from 7.84 to 1.33 for 240s generation.

Key takeaway

For machine learning engineers extending autoregressive video diffusion models to minute-level durations, TetherCache offers a critical, training-free solution. Its GRAB and TAME mechanisms directly mitigate accumulated context distribution shift, preventing visual artifacts and temporal drift. You should consider integrating these cache management principles to achieve stable, high-quality long-horizon video generation, especially when targeting 240s or longer outputs.

Key insights

TetherCache stabilizes long-form video generation by intelligently managing cache and aligning historical context.

Principles

• Gated recall can preserve diverse historical context.
• Memory editing reduces pollution from drifted features.

Method

TetherCache organizes cache into sink, memory, and recent regions. GRAB selects long-range memory frames using a gated score combining attention relevance and temporal diversity. TAME edits recalled memory tokens by aligning their statistics to a trusted context distribution.

In practice

• Combine attention-based relevance with temporal diversity for cache selection.
• Align statistics of recalled tokens to a trusted distribution.

Topics

• Autoregressive Video Generation
• Video Diffusion Models
• Cache Management
• Temporal Drift
• Gated Recall
• Memory Editing
• VBench-Long

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• CamDirector: Towards Long-Term Coherent Video Trajectory Editing — CamDirector enhances video trajectory editing with precise camera control and long-term consistency via hybrid warping and history-guided autoregressive diffusion.
• Real2SAM2Real: Generative 3D Caches as Complementary Context for Video Diffusion — Real2SAM2Real uses editable 3D caches from lifting models to provide robust geometric scaffolds for Video Diffusion Models, improving control and consistency.
• WorldKV: Efficient World Memory with World Retrieval and Compression — WorldKV uses retrieval and compression to maintain persistent world memory in video diffusion models efficiently.
• AnchorEdit: Maintaining Temporal Consistency in Multi-turn Image Editing via Causal Memory — AnchorEdit employs a causal memory mechanism and a three-stage training curriculum to ensure temporal consistency in multi-turn image editing.
• LongStream: Long-Sequence Streaming Autoregressive Visual Geometry — LongStream enables stable, metric-scale 3D reconstruction for ultra-long sequences by decoupling gauge and optimizing Transformer caches.
• Memory Caching: RNNs with Growing Memory — Memory Caching enhances RNNs by enabling memory growth with sequence length, bridging the performance gap with Transformers.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.