Pith. sign in

REVIEW 6 cited by

Training-free and Adaptive Sparse Attention for Efficient Long Video Generation

Not yet reviewed by Pith; the record is open.

Re-run · record.json Download PDF Read on arXiv ↗

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2502.21079 v1 pith:RHPWDFI2 submitted 2025-02-28 cs.CV

Yifei Xia , Suhan Ling , Fangcheng Fu , Yujie Wang , Huixia Li , Xuefeng Xiao , Bin Cui This is my paper · ORCID
classification cs.CV
keywords attentionditssparsepatternvideoacrossadaspablockified
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

Generating high-fidelity long videos with Diffusion Transformers (DiTs) is often hindered by significant latency, primarily due to the computational demands of attention mechanisms. For instance, generating an 8-second 720p video (110K tokens) with HunyuanVideo takes about 600 PFLOPs, with around 500 PFLOPs consumed by attention computations. To address this issue, we propose AdaSpa, the first Dynamic Pattern and Online Precise Search sparse attention method. Firstly, to realize the Dynamic Pattern, we introduce a blockified pattern to efficiently capture the hierarchical sparsity inherent in DiTs. This is based on our observation that sparse characteristics of DiTs exhibit hierarchical and blockified structures between and within different modalities. This blockified approach significantly reduces the complexity of attention computation while maintaining high fidelity in the generated videos. Secondly, to enable Online Precise Search, we propose the Fused LSE-Cached Search with Head-adaptive Hierarchical Block Sparse Attention. This method is motivated by our finding that DiTs' sparse pattern and LSE vary w.r.t. inputs, layers, and heads, but remain invariant across denoising steps. By leveraging this invariance across denoising steps, it adapts to the dynamic nature of DiTs and allows for precise, real-time identification of sparse indices with minimal overhead. AdaSpa is implemented as an adaptive, plug-and-play solution and can be integrated seamlessly with existing DiTs, requiring neither additional fine-tuning nor a dataset-dependent profiling. Extensive experiments validate that AdaSpa delivers substantial acceleration across various models while preserving video quality, establishing itself as a robust and scalable approach to efficient video generation.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 6 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Kaleido: Algorithm-Hardware Co-Design for Video Diffusion Transformers by Exploiting Latent Space Correlations

    cs.AR 2026-07 conditional novelty 6.0 of 10

    A channel-wise reuse algorithm plus a reconfigurable systolic accelerator skips redundant vDiT attention and MLP computation, achieving up to 5.9x speedup and 16x energy savings.

  2. SVG-EAR: Parameter-Free Linear Compensation for Sparse Video Generation via Error-aware Routing

    cs.CV 2026-03 conditional novelty 6.0 of 10

    Parameter-free centroid compensation plus error-aware block routing yields a better quality–density Pareto frontier for sparse attention in video DiTs than score-based sparsification.

  3. SnapGen++: Unleashing Diffusion Transformers for Efficient High-Fidelity Image Generation on Edge Devices

    cs.CV 2026-01 conditional novelty 6.0 of 10

    A compact elastic diffusion transformer with adaptive sparse attention and knowledge-guided distribution-matching distillation achieves 4-step 1K image generation on a phone in roughly 1.8 seconds.

  4. Trainable Log-linear Sparse Attention for Efficient Diffusion Transformers

    cs.CV 2025-12 conditional novelty 6.0 of 10

    LLSA uses hierarchical top-K selection plus coarse-token enrichment to make diffusion-transformer attention O(N log N), giving about 6x faster training at 256x256 pixel resolution with FID comparable to full attention.

  5. End-to-End Training for Autoregressive Video Diffusion via Self-Resampling

    cs.CV 2025-12 conditional novelty 6.0 of 10

    Resampling Forcing trains autoregressive video diffusion models on self-resampled degraded histories with a causal mask, achieving stable long-horizon generation without a teacher or discriminator.

  6. Timeripple: Accelerating vDiTs by Understanding the Spatio-Temporal Correlations in Latent Space

    cs.AR 2025-11 conditional novelty 6.0 of 10

    Timeripple cuts vDiT self-attention compute by up to 85% by reusing partial attention scores of spatially and temporally correlated tokens across channels, with VBench quality essentially unchanged.

Pith tools