Research map / Image generation / Autoregressive and latent generation
Autoregressive Image Generation with Randomized Parallel Decoding
Haopeng Li, Jinyue Yang, Guoqi Li, Huan Wang
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 |
|---|---|---|---|
| use block-wise causal attention at inference instead of causal (ARPG-L, 16 steps) (ablation) | ARPG-L · ImageNet 256x256 | FID: 3.51 -> 2.67 | Improved |
| use block-wise attention at 256 steps (ablation) | ARPG-L · ImageNet 256x256 | FID: 2.43 -> 2.44 | No clear effect |
| reduce decoding steps from 256 to 32 (ablation) | ARPG-L · ImageNet 256x256 | throughput (img/s): 17.45 -> 130.14 | Improved |
| reduce decoding steps from 256 to 32 (block-wise) (ablation) | ARPG-XL · ImageNet 256x256 | FID: 2.01 -> 1.93 | Improved |
| replace RoPE with cosine positional encoding (ablation) | ARPG-L · ImageNet 256 ablation | FID: 3.51 -> 3.61 | Hurt |
| remove shared KV (separate KV projections) (ablation) | ARPG-L · ImageNet 256 ablation | FID: 3.51 -> 3.46 | Mixed |
| make all layers second-pass decoders (0+24) (ablation) | ARPG-L · ImageNet 256 ablation | FID: 3.51 -> 4.57 | Hurt |
| use fewer pass-2 decoder layers (18+6) (ablation) | ARPG-L · ImageNet 256 ablation | throughput (img/s): 67.47 -> 62.35 | Hurt |
| add the two-pass decoder (vs none, standard AR with 24+0 layers) (ablation) | ARPG-L · ImageNet 256 ablation | throughput (img/s): 11.70 -> 67.47 | Improved |
| random generation order instead of raster order (ablation) | ARPG-L · ImageNet 256x256 | FID: 2.49 -> 2.44 | Improved |
| use spiral-out order instead of raster (ablation) | ARPG-L · ImageNet 256x256 | FID: 2.49 -> 4.11 | Hurt |
| ARPG-XXL vs LlamaGen-XXL | ARPG-XXL · ImageNet 256x256 | FID: 2.62 -> 1.83 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- RandAR-XXL — RandAR: Decoder-only Autoregressive Visual Generation in Random Orders (CVPR 2025) · better on 2 of 2 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
- RandAR: Decoder-only Autoregressive Visual Generation in Random Orders (CVPR 2025 · similarity 0.91)
- ZipAR: Parallel Autoregressive Image Generation through Spatial Locality (ICML 2025 · similarity 0.91)
- Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (NeurIPS 2025 · similarity 0.91)
- LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding (ICLR 2025 · similarity 0.89)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.89)
- AR-RAG: Autoregressive Retrieval Augmentation for Image Generation (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.
