Research map / Image generation / Autoregressive and latent generation
LazyVAR: Accelerating Visual Autoregressive Models via Scale-wise Token Pruning and Parallel Group Decoding
Rongge Mao, Chengqi Dong, S Kevin Zhou
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 |
|---|---|---|---|
| keep a moderate fraction of tokens at scales 9-10 instead of removing all (ablation) | LazyVAR on Infinity-2B · GenEval Overall | Overall: 0.658 -> 0.686 | Improved |
| increase retention ratios from [20%,10%] to [60%,30%] (ablation) | LazyVAR on Infinity-2B · GenEval / MJHQ-30K | Overall / time: 0.686 -> 0.685; 0.47 s -> 0.56 s | No clear effect |
| add Parallel Group Decoding (ablation) | LazyVAR on Infinity-2B · Inference speed | Speed-up: 2.03× -> 2.94× | Improved |
| group scale 9 as well ([9..13] instead of [10..13]) (ablation) | LazyVAR on Infinity-2B · GenEval / MJHQ-30K | Overall / FID*: 0.686 -> 0.678; FID* 1.46 -> 2.21 | Hurt |
| use Update Index criterion instead of FastVAR PTS (ablation) | LazyVAR on Infinity-2B · GenEval Overall | Overall: 0.665 -> 0.686 | Improved |
| add cumulative retained-token assignment to avoid updating same position across scales in a group (ablation) | LazyVAR on Infinity-2B · GenEval Overall | Overall: 0.681 -> 0.686 | Improved |
| LazyVAR instead of FastVAR on Infinity | Infinity-2B · MJHQ-30K | FID*: 2.03 -> 1.46 | Improved |
| raise FastVAR pruning ratio to match speed | FastVAR on Infinity-2B · GenEval / MJHQ-30K | Overall: 0.683 -> 0.662 | Hurt |
| apply LazyVAR to HART | HART-0.7B · GenEval Overall | Overall: 0.476 -> 0.468 | Hurt |
| apply LazyVAR to Infinity | Infinity-2B · HPSv2.1 average | HPS: 30.44 -> 29.83 | Hurt |
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 — Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis (CVPR 2025) · better on 3 of 4 shared results
- HART — HART: Efficient Visual Generation with Hybrid Autoregressive Transformer (ICLR 2025) · better on 3 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 0.91)
- Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient (CVPR 2025 · similarity 0.91)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.90)
- FlexVAR: Flexible Visual Autoregressive Modeling without Residual Prediction (NeurIPS 2025 · similarity 0.89)
- LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding (ICLR 2025 · similarity 0.89)
- Memory-Efficient Visual Autoregressive Modeling with Scale-Aware KV Cache Compression (NeurIPS 2025 · 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.
