Research map / Image generation / Autoregressive and latent generation
Context-Aware Autoregressive Image Generation for Emerging Reasoning Properties
Jixuan Ying, Haoyu Liu, Timing Yang, Tingyu Zhu, Predrag Neskovic, Anand Bhattad, Cihang Xie, Zeyu Zheng, Alan Yuille, Feng Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a context-aware generation-order approach for autoregressive image generation to better capture structural dependencies and reasoning properties, as stated in the abstract.
Problem
Autoregressive image generation uses predefined or heuristic generation orders (next-patch, random-order, one-step per scale), which can ignore important structural dependencies in images and limit structural robustness and reasoning.
Approach
Makes the generation order context-aware rather than fixed or heuristic, so the autoregressive model follows image structural dependencies; the abstract does not give further details of the mechanism.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace random order with SAR context-aware order at inference | MaskGen-XL, Infinity-2B, Infinity-8B · Complex long-prompt benchmark | average score (1-5): 2.42 -> 2.92 (MaskGen-XL); 2.92 -> 3.42 (Infinity-8B) | Improved |
| replace random order with SAR order | MaskGen-L, MaskGen-XL, Infinity-2B, Infinity-8B · GenEval | Overall pass rate (%): 56.22 -> 59.99 (MaskGen-XL) | Improved |
| replace random order with SAR order | MaskGen and Infinity models · DPG-Bench | Overall (%): 86.09 -> 86.27 (Infinity-8B); 70.11 -> 70.51 (MaskGen-L) | Improved |
| replace random order with SAR order | MaskGen and Infinity models · MJHQ-30K | FID: 10.867 -> 10.970 (Infinity-2B); 9.332 -> 9.317 (Infinity-8B) | No clear effect |
| use SAR order instead of sequential order (ablation) | MaskGen-XL · GenEval | GenEval score: 0.57 -> 0.58 | Improved |
| use SAR order instead of inside-out order (ablation) | Infinity-8B · GenEval | GenEval score: 0.808 -> 0.815 | Improved |
| use sequential order instead of random (ablation) | Infinity-8B · GenEval | GenEval score: 0.804 -> 0.799 | No clear effect |
| replace confidence-based position selection with SAR in MaskGIT-Base without retraining (ablation) | MaskGIT-Base · ImageNet 256x256 class-conditional | FID: 6.48 -> 5.95 | Improved |
| add SAR ranking on top of VAR last-scale decoding (ablation) | VAR · VAR inference latency (H100, batch 1) | latency | No clear effect |
Closest papers by meaning
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.92)
- GoT-R1: Unleashing Reasoning Capability of Autoregressive Visual Generation with Reinforcement Learning (ICLR 2026 · similarity 0.91)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.90)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.90)
- OmniGen-AR: AutoRegressive Any-to-Image Generation (NeurIPS 2025 · similarity 0.90)
- Parallelized Autoregressive Visual Generation (CVPR 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.
