Research map / Image generation / Autoregressive and latent generation
Next Patch Prediction for AutoRegressive Visual Generation
Yatian Pang, Peng Jin, Shuo Yang, Bin Zhu, Bin Lin, Chaoran Feng, Zhenyu Tang, Liuhan Chen, Francis E. H. Tay, Ser-Nam Lim, Harry Yang, Li Yuan
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 |
|---|---|---|---|
| train with next patch prediction (2x2 patches then 1x1) instead of next token prediction | LlamaGen-B · ImageNet 256x256 | FID: 5.46 -> 4.47 | Improved |
| train with next patch prediction | LlamaGen-L · ImageNet 256x256 | FID: 3.80 -> 2.76 | Improved |
| train with next patch prediction | LlamaGen-XL · ImageNet 256x256 | FID: 3.39 -> 2.65 | Improved |
| train with next patch prediction | LlamaGen-XXL · ImageNet 256x256 | FID: 3.10 -> 2.54 | Improved |
| train with next patch prediction (training compute) | LlamaGen-L · ImageNet 256x256 | training GFLOPs: 83.54 -> 47.95 | Improved |
| enlarge first-stage patch size from 2x2 to 4x4 (ablation) | NPP-B · ImageNet 256x256 | FID: 4.47 -> 4.92 | Hurt |
| enlarge first-stage patch size from 2x2 to 4x4 (ablation) | NPP-L · ImageNet 256x256 | FID: 2.76 -> 2.89 | Hurt |
| raise segment factor lambda from 1/2 to 4/5 (ablation) | NPP-L · ImageNet 256x256 | FID: 2.76 -> 2.92 | Hurt |
| raise segment factor lambda from 1/2 to 4/5 (cost) (ablation) | NPP-L · ImageNet 256x256 | training cost: 0.625x -> 0.4x | Improved |
| use more patch levels (#L=2 to 4) (ablation) | NPP-L · ImageNet 256x256 | FID: 2.76 -> 2.84 | Hurt |
| use more patch levels (#L=2 to 4) (ablation) | NPP-B · ImageNet 256x256 | FID: 4.47 -> 4.68 | Hurt |
| use NPP-XL instead of LlamaGen-XL-384 | NPP-XL · ImageNet 256x256 | FID: 2.62 vs 2.65 | No clear effect |
Closest papers by meaning
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.92)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.91)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.91)
- FlexVAR: Flexible Visual Autoregressive Modeling without Residual Prediction (NeurIPS 2025 · similarity 0.91)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.90)
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive 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.
