Research map / Image generation / Autoregressive and latent generation
SCAN: Self-Calibrated AutoregressioN for High-Quality Visual Generation
Zhanzhou Feng, Qingpei Guo, Jingdong Chen, Feng Gao, Ming Yang, Shiliang Zhang
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 30 refinement iterations at inference (ablation) | SCAN-XL · ImageNet 256x256 | FID: 2.52 -> 2.10 | Improved |
| add 30 refinement iterations at inference (ablation) | SCAN-L · ImageNet 256x256 | FID: 3.05 -> 2.51 | Improved |
| add 30 refinement iterations at inference (ablation) | SCAN-XXL · ImageNet 256x256 | FID: 2.24 -> 1.95 | Improved |
| increase refinement iterations from 30 to 60 (ablation) | SCAN-XL · ImageNet 256x256 | FID: 2.10 -> 2.03 (latency 3.56 -> 5.11) | No clear effect |
| train with only random-noise negatives (ablation) | SCAN-XL · ImageNet 256x256 | FID (initial generation): 2.52 -> 3.27 | Hurt |
| train with only location-permuted negatives (ablation) | SCAN-XL · ImageNet 256x256 | FID (initial generation): 2.52 -> 2.87 | Hurt |
| train with only teacher-forcing negatives (ablation) | SCAN-XL · ImageNet 256x256 | FID (with refinement): 2.10 -> 2.31 | Hurt |
| combine three negative sample types (ablation) | SCAN-XL · ImageNet 256x256 | FID (with refinement): 2.31 -> 2.10 (vs teacher forcing only) | Improved |
| use SCAN-XL instead of LlamaGen-XL | SCAN-XL · ImageNet 256x256 | latency (s): 19.76 -> 3.56 | 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-XL — RandAR: Decoder-only Autoregressive Visual Generation in Random Orders (CVPR 2025) · better on 3 of 3 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
- Generation then Reconstruction: Accelerating Masked Autoregressive Models via Two-Stage Sampling (ICLR 2026 · similarity 0.89)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.88)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.88)
- EditAR: Unified Conditional Generation with Autoregressive Models (CVPR 2025 · similarity 0.87)
- RandAR: Decoder-only Autoregressive Visual Generation in Random Orders (CVPR 2025 · similarity 0.86)
- HART: Efficient Visual Generation with Hybrid Autoregressive Transformer (ICLR 2025 · similarity 0.86)
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.
