{"paper":{"title":"HASTE: Training-Free Video Diffusion Acceleration via Head-Wise Adaptive Sparse Attention","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Head-wise adaptive sparse attention accelerates pretrained video diffusion models up to 1.93 times without retraining by reusing temporal masks and calibrating sparsity per head.","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Fei Chao, Jing Xu, Rongrong Ji, Xiawu Zheng, Xuzhe Zheng, Yuexiao Ma","submitted_at":"2026-05-14T07:57:55Z","abstract_excerpt":"Diffusion-based video generation has advanced substantially in visual fidelity and temporal coherence, but practical deployment remains limited by the quadratic complexity of full attention. Training-free sparse attention is attractive because it accelerates pretrained models without retraining, yet existing online top-$p$ sparse attention still spends non-negligible cost on mask prediction and applies shared thresholds despite strong head-level heterogeneity. We show that these two overlooked factors limit the practical speed-quality trade-off of training-free sparse attention in Video DiTs. "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"On Wan2.1-1.3B and Wan2.1-14B, our method consistently improves XAttention and SVG2, achieving up to 1.93 times speedup at 720P while maintaining competitive video quality and similarity metrics.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That measured model-output error under a global sparsity budget reliably predicts perceptual video quality across heads and that temporal query-key drift is stable enough for safe mask reuse without visible artifacts.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"HASTE delivers up to 1.93x speedup on Wan2.1 video DiTs via head-wise adaptive sparse attention using temporal mask reuse and error-guided per-head calibration while preserving video quality.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Head-wise adaptive sparse attention accelerates pretrained video diffusion models up to 1.93 times without retraining by reusing temporal masks and calibrating sparsity per head.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"12881d9dd2943bf8b402cebdcd01c9ee463c129c904b2ad940caa4efa71b338f"},"source":{"id":"2605.14513","kind":"arxiv","version":1},"verdict":{"id":"743a148e-dd17-4d7b-abeb-b7248690224f","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T05:14:18.704869Z","strongest_claim":"On Wan2.1-1.3B and Wan2.1-14B, our method consistently improves XAttention and SVG2, achieving up to 1.93 times speedup at 720P while maintaining competitive video quality and similarity metrics.","one_line_summary":"HASTE delivers up to 1.93x speedup on Wan2.1 video DiTs via head-wise adaptive sparse attention using temporal mask reuse and error-guided per-head calibration while preserving video quality.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That measured model-output error under a global sparsity budget reliably predicts perceptual video quality across heads and that temporal query-key drift is stable enough for safe mask reuse without visible artifacts.","pith_extraction_headline":"Head-wise adaptive sparse attention accelerates pretrained video diffusion models up to 1.93 times without retraining by reusing temporal masks and calibrating sparsity per head."},"references":{"count":55,"sample":[{"doi":"","year":2025,"title":"Dicache: Let diffusion model determine its own cache","work_id":"c2389322-8839-490a-8d61-34832e66ebb6","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Rainfusion: Adaptive video generation acceleration via multi-dimensional visual redundancy","work_id":"a6bfc322-49cb-4f3c-bf04-19fcfb1e39b5","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"RainFusion2.0: Temporal-Spatial Awareness and Hardware-Efficient Block-wise Sparse Attention","work_id":"179d3710-8907-440f-8695-8d1a532bfe14","ref_index":3,"cited_arxiv_id":"2512.24086","is_internal_anchor":true},{"doi":"","year":2025,"title":"Sparse-vDiT: Unleashing the Power of Sparse Attention to Accelerate Video Diffusion Transformers, June 2025","work_id":"8ff929a3-a68f-4193-909e-1b19fbdb4f57","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Hicache: Training-free acceleration of diffusion models via hermite polynomial- based feature caching.arXiv preprint arXiv:2508.16984","work_id":"43ebda46-bf78-416c-9a3e-a372d3d1b9e7","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":55,"snapshot_sha256":"de4fd7557ab4c60644eb972a4e21a505b7104439c8879b76ffea616e27d90624","internal_anchors":11},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}