HiAR: Efficient Autoregressive Long Video Generation via Hierarchical Denoising

2026-03-09 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Expert, quick

Summary

HiAR is a novel hierarchical denoising framework designed for efficient autoregressive long video generation, addressing the challenge of temporal continuity and quality degradation in existing methods. Unlike conventional approaches that condition on highly denoised contexts, HiAR proposes conditioning on context at the same noise level as the current block, which sufficiently maintains temporal consistency while mitigating error propagation. This framework reverses the typical sequential block generation order, instead performing causal generation across all blocks at each denoising step. This design enables pipelined parallel inference, achieving a 1.8x wall-clock speedup in a 4-step setting. Additionally, HiAR incorporates a forward-KL regularizer in bidirectional-attention mode to counteract a low-motion shortcut amplified by self-rollout distillation, thereby preserving motion diversity. On the VBench benchmark (20s generation), HiAR achieved the best overall score and the lowest temporal drift.

Key takeaway

For research scientists developing long video generation models, HiAR's hierarchical denoising approach offers a significant advancement. You should consider adopting its strategy of conditioning on same-noise-level contexts and its forward-KL regularization to improve temporal consistency, mitigate error accumulation, and enhance motion diversity in your autoregressive diffusion models. This could lead to more efficient and higher-quality long video outputs.

Key insights

Conditioning on same-noise-level context in autoregressive diffusion improves temporal consistency and mitigates error propagation.

Principles

• Highly clean context is unnecessary for temporal continuity.
• Same-noise-level conditioning provides sufficient temporal signal.

Method

HiAR reverses conventional generation order, performing causal generation across all blocks at each denoising step, conditioning on same-noise-level context. It uses a forward-KL regularizer to preserve motion diversity.

In practice

• Achieves 1.8x wall-clock speedup via pipelined parallel inference.
• Reduces temporal drift in long video generation.

Topics

• Autoregressive Diffusion
• Long Video Generation
• Hierarchical Denoising
• Temporal Consistency
• Motion Diversity

Best for: Research Scientist, AI Researcher, AI Scientist, Computer Vision Engineer

Related on AIssential

• Consistency-Preserving Diverse Video Generation — A new flow-matching framework improves video diversity and temporal consistency using latent-space gradient regulation.
• Normalizing Flows with Iterative Denoising — iTARFlow advances Normalizing Flows with a likelihood-based objective and diffusion-inspired iterative denoising for image generation.
• Future Forcing: Future-aware Training-free KV Cache Policy for Autoregressive Video Generation — Future Forcing uses stable pre-RoPE query distributions to enable training-free, future-aware KV cache management for AR video generation.
• Diffusion Models for Adaptive Sequential Data Generation — A sequential forward-backward diffusion framework generates adaptive time series by conditioning noise injection and removal on historical data.
• Veda: Scalable Video Diffusion via Distilled Sparse Attention — Generation quality in sparse attention depends on mask alignment with full attention's tile-wise geometry, not just sparsity ratio.
• I made a beginner-friendly visual explanation of how Stable Diffusion works (feedback welcome) — Stable Diffusion generates images by iteratively denoising random noise, guided by text embeddings and learned visual structures.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.