Research map / Video / Video generation
QuantSparse: Comprehensively Compressing Video Diffusion Transformer with Model Quantization and Attention Sparsification
Weilun Feng, Chuanguang Yang, Haotong Qin, Mingqiang Wu, Yuqi Li, Xiangqi Li, Zhulin An, Libo Huang, Yulun Zhang, Michele Magno, Yongjun Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add global attention distillation (ablation) | QuantSparse · Wan2.1-14B W4A8 | PSNR: 14.35 -> 16.01 | Improved |
| add local salient distillation (ablation) | QuantSparse · Wan2.1-14B W4A8 | PSNR: 14.35 -> 16.82 | Improved |
| combine global and local distillation (MSAD) (ablation) | QuantSparse · Wan2.1-14B W4A8 | PSNR: 14.35 -> 18.72 | Improved |
| select top-k salient queries instead of random queries for local distillation (ablation) | QuantSparse · Wan2.1-14B W4A8 | PSNR: 15.49 -> 16.82 | Improved |
| add first-order residual reparameterization (ablation) | QuantSparse · Wan2.1-14B W4A8 | PSNR: 14.16 -> 17.08 | Improved |
| use second-order residual instead of first-order (ablation) | QuantSparse · Wan2.1-14B W4A8 | VQA: 70.82 -> 89.73 | Improved |
| add SVD projection to second-order residual (SSAR) (ablation) | QuantSparse · Wan2.1-14B W4A8 | VQA: 89.73 -> 91.98 | Improved |
| shorten cache refresh interval from 5 to 3 (ablation) | QuantSparse · Wan2.1-14B W4A8 | PSNR: 18.22 -> 18.86 | Mixed |
| lengthen cache refresh interval from 5 to 6 (ablation) | QuantSparse · Wan2.1-14B W4A8 | PSNR: 18.22 -> 17.72 | Hurt |
| naively combine quantization with sparse attention (Q-VDiT+SVG) | Q-VDiT · HunyuanVideo W4A8 25% density | PSNR: 16.85 -> 16.71 | No clear effect |
| combine quantization and sparsification (attention shift) (ablation) | QuaRot + SVG · Wan2.1-1.3B W4A8 | attention shift (MSE): 0.216 and 0.134 separately -> 0.685 combined | Hurt |
| lower attention density from 15% to 10% (ablation) | QuantSparse · Wan2.1-14B W4A8 | PSNR: 18.22 -> 17.73 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Q-VDiT — Q-VDiT: Towards Accurate Quantization and Distillation of Video-Generation Diffusion Transformers (ICML 2025) · better on 7 of 7 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- $\text{S}^2$Q-VDiT: Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation (NeurIPS 2025 · similarity 0.93)
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.93)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.91)
- DiTFastAttn: Attention Compression for Diffusion Transformer Models (NeurIPS 2024 · similarity 0.90)
- Q-DiT: Accurate Post-Training Quantization for Diffusion Transformers (CVPR 2025 · similarity 0.90)
- Re-ttention: Ultra Sparse Visual Generation via Attention Statistical Reshape (NeurIPS 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
