headcast-attention-heads-video-generation

name	headcast-attention-heads-video-generation
version	1.0.0
description	HeadCast methodology for efficient autoregressive video generation through training-free attention head classification and KV cache optimization.
author	Jinliang Shen, Lianghao Su, Zheming Li, Kang He, ZiLiang Lai et al.
license	MIT
arxiv_id	2607.20125v1
tags	["video-generation","attention","efficiency","kv-cache","autoregressive"]

HeadCast: Casting Attention Heads for Efficient Autoregressive Video Generation

Overview

HeadCast is a training-free, plug-and-play acceleration framework for autoregressive (AR) video diffusion models that addresses the continuously growing Key-Value (KV) cache problem during inference.

Key Components

Attention Head Classification

After a short warm-up period, HeadCast performs one-time classification at the maximum-noise step to sort every attention head into one of four archetypes:

• Sink Heads: Handle attention sink phenomena
• Dummy Heads: Can be safely pruned or approximated
• Spatial Heads: Focus on spatial relationships within frames
• Global Heads: Preserve long-range temporal consistency

Head-Specific Pathways

• Restructures the monolithic KV cache into head-specific pathways
• Retains Global heads to maintain temporal consistency that aggressive eviction would destroy
• Spatial pathway operates on a fixed-size grid, providing resolution-dependent savings

Implementation Guidelines

Integration Workflow

1. Load pre-trained AR video diffusion model
2. Perform short warm-up inference to observe head behaviors
3. Execute one-time classification at maximum-noise step
4. Restructure KV cache into head-specific pathways
5. Continue inference with optimized attention computation

Performance Characteristics

• Acceleration improves with higher resolution (1.62x at 720P, 1.95x at 1080P)
• Maintains VBench quality comparable to full attention
• Largely eliminates inter-frame flickering
• No re-training required

Use Cases

• High-resolution autoregressive video generation
• Streaming video synthesis
• Real-time video diffusion applications
• Memory-constrained inference scenarios

Activation Keywords

HeadCast, video generation, attention heads, KV cache optimization, autoregressive diffusion, inference acceleration

References

• arXiv: 2607.20125v1
• Authors: Jinliang Shen, Lianghao Su, Zheming Li, Kang He, ZiLiang Lai et al.
• Published: July 22, 2026
• Code: https://github.com/sjlgaga/HeadCast