Research map / Image generation / Autoregressive and latent generation
Randomized Autoregressive Visual Generation
Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, Liang-Chieh Chen
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 |
|---|---|---|---|
| anneal randomness (start 200, end 300) versus pure raster order (ablation) | RAR-L · ImageNet-256 (RAR-L) | FID: 3.08 -> 2.18 | Improved |
| pure random order training versus raster order (ablation) | RAR-L · ImageNet-256 (RAR-L) | FID: 3.08 -> 3.01 | No clear effect |
| anneal randomness (start 200, end 300) versus pure random order (ablation) | RAR-L · ImageNet-256 (RAR-L) | FID: 3.01 -> 2.18 | Improved |
| replace first 100 epochs of raster order with random permutation (start 100, end 100) (ablation) | RAR-L · ImageNet-256 (RAR-L) | FID: 3.08 -> 2.48 | Improved |
| start annealing later (start 300, end 400) instead of 200-300 (ablation) | RAR-L · ImageNet-256 (RAR-L) | FID: 2.18 -> 2.74 | Hurt |
| start annealing at 100 and end at 200 instead of 200-300 (ablation) | RAR-L · ImageNet-256 (RAR-L) | FID: 2.18 -> 2.28 | Hurt |
| use z-curve final scan order instead of row-major (ablation) | RAR-L · ImageNet-256 (RAR-L) | FID: 2.18 -> 2.29 | Hurt |
| use spiral in final scan order instead of row-major (ablation) | RAR-L · ImageNet-256 (RAR-L) | FID: 2.18 -> 2.50 | Hurt |
| use alternate final scan order instead of row-major (ablation) | RAR-L · ImageNet-256 (RAR-L) | FID: 2.18 -> 2.48 | Hurt |
| scale model from RAR-B to RAR-XXL | RAR · ImageNet-256 | FID: 1.95 -> 1.48 | Improved |
| use RAR-XL versus MaskBit at similar FID | RAR-XL · ImageNet-256 sampling speed | images/sec: 0.7 -> 8.3 | Improved |
| use RAR-XXL versus LlamaGen-3B-384 | RAR-XXL · ImageNet-256 | FID: 2.18 -> 1.48 | Improved |
Closest papers by meaning
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.88)
- RandAR: Decoder-only Autoregressive Visual Generation in Random Orders (CVPR 2025 · similarity 0.88)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.87)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.87)
- Autoregressive Image Generation without Vector Quantization (NeurIPS 2024 · similarity 0.87)
- HART: Efficient Visual Generation with Hybrid Autoregressive Transformer (ICLR 2025 · similarity 0.87)
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.
