Research map / Image generation / Autoregressive and latent generation
Neighboring Autoregressive Modeling for Efficient Visual Generation
Yefei He, Yuanyu He, Shaoxuan He, Feng Chen, Hong Zhou, Kaipeng Zhang, Bohan Zhuang
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 dimension-oriented decoding heads (ablation) | NAR-L · ImageNet 256x256 | FID: 66.31 -> 3.06 | Improved |
| mix logits of horizontal and vertical heads on overlapped tokens (ablation) | NAR-L · ImageNet 256x256 | FID: 75.91 -> 3.06 (horizontal only), 25.73 -> 3.06 (vertical only) | Improved |
| mix logits of horizontal and vertical heads on overlapped tokens (ablation) | NAR-L · ImageNet 256x256 | IS: 22.36 -> 263.9 (horizontal only), 96.08 -> 263.9 (vertical only) | Improved |
| replace next-token prediction with next-neighbor prediction | LlamaGen-L · ImageNet 256x256 | FID and steps: 3.80 -> 3.06, 256 -> 31 steps | Improved |
| NAR-L (372M) vs LlamaGen-XXL (1.4B) | LlamaGen-XXL · ImageNet 256x256 | throughput: 14.1 -> 195.4 img/s | Improved |
| replace multi-scale next-scale prediction with NAR | VAR-d16 · ImageNet 256x256 | FID and throughput: 3.30 -> 3.27, 129.3 -> 248.5 img/s | Improved |
| replace PAR-4X parallel decoding with NAR | PAR-XL-4X · ImageNet 256x256 | FID: 3.50 -> 2.70 | Improved |
| use NAR instead of LARP-L-Long autoregression | LARP-L-Long · UCF-101 | FVD and latency: reducing FVD by 5.8 and generation latency by 97.5% | Improved |
| use NAR-XL instead of PAR-XL parallel decoding | PAR-XL · UCF-101 | FVD: 23~32.3 FVD reduction | Improved |
| increase PAR parallelism from 1x to 16x | PAR-XL · UCF-101 | FVD: 94.1 -> 99.5 -> 103.4 | Hurt |
| scale NAR-L to deeper NAR-LP | NAR-L · UCF-101 | FVD: 96.2 -> 71.1 | Improved |
| swap LlamaGen-XL for NAR-XL in text-to-image | LlamaGen-XL · GenEval | overall score: 0.32 -> 0.43 | Improved |
Closest papers by meaning
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.93)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.92)
- ZipAR: Parallel Autoregressive Image Generation through Spatial Locality (ICML 2025 · similarity 0.91)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.91)
- Autoregressive Video Generation without Vector Quantization (ICLR 2025 · similarity 0.90)
- OmniGen-AR: AutoRegressive Any-to-Image Generation (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.
