Research map / Image generation / Autoregressive and latent generation
ARVAR: Accelerating Visual Autoregressive Model via Attention Retrospect
Jiedong Zhuang, Lu Lu, Ming Dai, Jian Chen, Qiang Liu, Haoji Hu
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 |
|---|---|---|---|
| replace frequency-based token selection with attention-saliency (self-attn) selection (ablation) | ARVAR on HART · MJHQ30K | FID: 28.19 -> 27.67 | Improved |
| use random token pruning instead of saliency (ablation) | ARVAR on HART · MJHQ30K | FID: 27.67 -> 31.98 | Hurt |
| use self-attention part rather than full attention for saliency (ablation) | ARVAR on HART · MJHQ30K | FID: 28.23 -> 27.67 | Improved |
| add Attention Retrospect alone (ablation) | HART · MJHQ30K | FID / speedup: 30.75 -> 27.56, 1.46x | Improved |
| add Branch Halving alone (ablation) | HART · MJHQ30K | FID / speedup: 30.75 -> 28.01, 1.43x | Improved |
| combine AR and BH (ablation) | HART · MJHQ30K | speedup: 1.46x -> 1.56x speedup, FID 27.56 -> 27.67, SSIM 0.7232 -> 0.6904 | Mixed |
| ARVAR instead of FastVAR | HART · MJHQ30K | SSIM / FID: SSIM 0.6689 -> 0.6904, FID 28.19 -> 27.67 | Improved |
| ARVAR instead of FastVAR | Infinity-2B · MJHQ30K | PSNR: 20.47 -> 22.73 | Improved |
| apply ARVAR to HART on HPSv2.1 | HART · HPSv2.1 | HPSv2.1: 28.72 -> 27.67 | Hurt |
| apply ARVAR to Infinity-2B | Infinity-2B · GenEval | overall: 0.68 -> 0.69 | No clear effect |
| prune at later step with higher ratio (75% at S14 vs 50% at S13) (ablation) | ARVAR on HART · MJHQ30K | FID / speedup: 28.12 -> 27.67 FID | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Infinity-2B — Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis (CVPR 2025) · better on 2 of 3 shared results
- HART — HART: Efficient Visual Generation with Hybrid Autoregressive Transformer (ICLR 2025) · better on 2 of 3 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
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.93)
- SparVAR: Exploring Sparsity in Visual AutoRegressive Modeling for Training-Free Acceleration (CVPR 2026 · similarity 0.92)
- FasterVAR: Plug-and-Play Acceleration for Visual Autoregressive Models (ICML 2026 · similarity 0.91)
- FlexVAR: Flexible Visual Autoregressive Modeling without Residual Prediction (NeurIPS 2025 · similarity 0.91)
- OmniGen-AR: AutoRegressive Any-to-Image Generation (NeurIPS 2025 · similarity 0.90)
- ExtraVAR: Stage-Aware RoPE Remapping for Resolution Extrapolation in Visual Autoregressive Models (NeurIPS 2026 · similarity 0.89)
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.
