Research map / Image generation / Autoregressive and latent generation
reAR: Rethinking Visual Autoregressive Models via Token-wise Consistency Regularization
Qiyuan He, Yicong Li, Haotian Ye, Jinghao Wang, Xinyao Liao, Pheng-Ann Heng, Stefano Ermon, James Zou, Angela Yao
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 reAR (noisy context + codebook embedding regularization) | raster AR-L with MaskGIT VQGAN · ImageNet-1K 256x256 | gFID: 3.02 -> 1.86 | Improved |
| add reAR to non-standard tokenizers | TiTok and AliTok AR · ImageNet-1K 256x256 | gFID: TiTok 4.45 -> 4.01; AliTok 1.50 -> 1.42 | Improved |
| add decoding embedding regularization at layer 15 (predict next-token embedding) (ablation) | vanilla AR (reAR-S, 80 ep) · ImageNet-256 no CFG | FID: 21.32 -> 20.03 | Improved |
| place decoding regularization at early layer 10 (ablation) | vanilla AR (reAR-S, 80 ep) · ImageNet-256 no CFG | FID: 21.32 -> 21.29 | No clear effect |
| tie codebook embeddings to the AR model directly (ablation) | vanilla AR (reAR-S, 80 ep) · ImageNet-256 no CFG | FID: 21.32 -> 21.08 | No clear effect |
| add encoding regularization at layer 0 on top of DE@15 (ablation) | reAR-S (80 ep) · ImageNet-256 no CFG | FID: 20.03 -> 19.72 | Improved |
| apply encoding regularization at layer 5 instead of layer 0 (ablation) | reAR-S (80 ep) · ImageNet-256 no CFG | FID: 19.72 -> 21.36 | Hurt |
| vary regularization weight lambda (0.5, 1, 1.5) (ablation) | reAR-S (80 ep) · ImageNet-256 no CFG | FID: 19.79 / 19.72 / 19.74 | No clear effect |
| add fixed noisy context (eps=0.25) to embedding regularization (ablation) | reAR-S (400 ep) · ImageNet-256 | FID: 2.12 -> 2.08 | Improved |
| raise fixed noise level to 0.5 (ablation) | reAR-S (400 ep) · ImageNet-256 | FID: 2.12 -> 3.15 | Hurt |
| anneal the maximum noise level over training (truncated linear) (ablation) | reAR-S (400 ep) · ImageNet-256 | FID: 2.05 -> 2.00 | Improved |
| add embedding regularization to noisy context regularization (ablation) | reAR-S (400 ep) · ImageNet-256 | FID: 2.18 -> 2.00 | Improved |
Closest papers by meaning
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · similarity 0.92)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.91)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.91)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.90)
- OmniGen-AR: AutoRegressive Any-to-Image Generation (NeurIPS 2025 · similarity 0.90)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 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.
