Research map / Image generation / Autoregressive and latent generation
SparVAR: Exploring Sparsity in Visual AutoRegressive Modeling for Training-Free Acceleration
Zekun Li, Ning Wang, Tongxin Bai, Changwang Mei, Peisong Wang, Shuang Qiu, Jian Cheng
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 early-scale attention sink to CSLA sparse mask (ablation) | CSLA on Infinity-8B · GenEval/PSNR | PSNR: 23.541 -> 25.897 | Improved |
| add cached dense output O_cache to CS4A (ablation) | CS4A on Infinity-8B · Table 5 | PSNR: 25.665 -> 26.359 | Improved |
| add CSLA block-wise local sparse attention on top of CS4A (ablation) | SparVAR on Infinity-8B · Table 5 | PSNR: 26.359 -> 29.481; speedup 1.43x -> 1.57x | Improved |
| use block size 128 instead of 64 in CSLA kernel (ablation) | CSLA FlexAttention kernel · last-scale attention forward | speedup vs FlashAttention2: 3.60x -> 5.61x | Improved |
| use smaller local windows ([3,3,3] vs [3,5,7]) in CSLA (ablation) | CSLA · Infinity-8B | latency/PSNR: latency 0.5346 -> 0.4480 ms; PSNR 25.897 -> 25.837 | Mixed |
| use a larger sink range (scale <=8 vs <=5) (ablation) | CSLA · Infinity-8B | PSNR / latency: PSNR 25.897 -> 27.582; latency 0.5346 -> 0.7729 ms | Mixed |
| choose a mid sparse decision scale (10) rather than early scales (6-7) (ablation) | CS4A · Infinity-8B | generation quality | Improved |
| replace scale skipping (FastVAR) with SparVAR sparse attention without skipping | Infinity-8B · Infinity-8B GenEval 1024 | PSNR vs baseline: 17.403 -> 29.481 | Improved |
| combine SparVAR with skip-last-2-scales | Infinity-8B · GenEval | speedup: 1.57x -> 2.28x | Improved |
| SparVAR vs unaccelerated baseline on semantic quality | Infinity-8B · GenEval overall | GenEval: 0.798 -> 0.796 | No clear effect |
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.
- Infinity-8B — Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis (CVPR 2025) · better on 4 of 5 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
- ARVAR: Accelerating Visual Autoregressive Model via Attention Retrospect (ECCV 2026 · similarity 0.92)
- Frequency-Aware Autoregressive Modeling for Efficient High-Resolution Image Synthesis (ICCV 2025 · similarity 0.91)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.91)
- FlexVAR: Flexible Visual Autoregressive Modeling without Residual Prediction (NeurIPS 2025 · similarity 0.90)
- Memory-Efficient Visual Autoregressive Modeling with Scale-Aware KV Cache Compression (NeurIPS 2025 · similarity 0.88)
- HMAR: Efficient Hierarchical Masked Auto-Regressive Image Generation (CVPR 2025 · similarity 0.88)
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.
