Research map / Video / Video generation

iCATS: Fast Video Generation via Interaction-Aware Sparse Attention and Timestep-Adaptive Sparsity

NeurIPS 2026 · Video › Video generation · approach: Efficient video diffusion models

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes iCATS, a training-free sparse attention method for DiT video generation that clusters tokens by query-key interactions via a quadratic form, uses an SNR-guided timestep-adaptive top-p schedule, and a tail-merging strategy for GPU kernel efficiency.

Problem

Training-free sparse attention for Diffusion Transformer video generation trades off efficiency against quality; existing cluster-based importance estimation uses feature-space clustering that misaligns with the query-key interaction error, fixed top-p sparsity ignores timestep-varying error toleran

Approach

Cluster keys by their query-key dot-product interaction error, which reduces to a d×d quadratic form (M_Q, M_K) computed from centroids and sizes; set top-p per timestep via an inverted sigmoid of log-SNR; and merge tail blocks of query clusters with Hamming-nearest compatible clusters to reduce padding.

Key results

• On HunyuanVideo-T2V-13B, iCATS-Base achieves 2.03x speedup with 31.017 dB PSNR vs. 26.247 dB for SVG2 at the same speed.
• On Wan2.1-T2V-14B, iCATS-Base achieves 1.55x speedup with 29.301 dB PSNR vs. 25.606 dB for SVG2.
• In human preference (220 votes, 11 participants), iCATS-Base won 54.1% vs. 18.6% for SVG2 on Wan2.1-T2V-14B.

Surprising findings

• Replacing the M_Q/M_K approximation with exact Q^T Q / K^T K gave the same token-level overlap (0.912) while increasing denoising time (1257 s vs 1219 s).
• Removing the sparsity-matched cluster-based approach is not needed for quality; the tail-merging step slightly reduced sparsity yet improved speed.

Limitations

• Clustering procedure is not optimized for multi-GPU parallel execution.
• Token-level sparse masking kernel is not yet fully optimized.
• Comparisons limited to SVG and SVG2 as baselines.

Experimental setup

Datasets: Penguin Benchmark prompts VBench prompts

Benchmarks: VBench Penguin Benchmark

Metrics: PSNR SSIM LPIPS VBench speedup sparsity top-p overlap output MSE

Baselines: SVG SVG2

Builds on: SVG2 Flash-Kmeans FlashInfer SVG

Closest papers by meaning

• Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.90)
• Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.89)
• Sparse VideoGen2: Accelerate Video Generation with Sparse Attention via Semantic-Aware Permutation (NeurIPS 2025 · similarity 0.89)
• Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.89)
• MonarchRT: Efficient Attention for Real-Time Video Generation (NeurIPS 2026 · similarity 0.89)
• Parameterized Stripe Attention for Efficient Video Generation (NeurIPS 2026 · similarity 0.89)