Research map / Image generation / Fast diffusion sampling
Revisiting Autoregressive Models for Generative Image Classification
Ilia Sudakov, Artem Babenko, Dmitry Baranchuk
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 |
|---|---|---|---|
| marginalize over multiple token orders (K=20) instead of raster order (ablation) | RandAR L/16 · ImageNet-Val | top-1 accuracy: 0.701 -> 0.780 | Improved |
| single random order (K=1) instead of raster order (ablation) | RandAR · ImageNet-Val | top-1 accuracy: 0.712 vs raster-order better at K=1 | Hurt |
| increase number of orders K from 1 to 20 (lower bound) (ablation) | RandAR XL/16 · ImageNet-Val | top-1 accuracy: 0.712 -> 0.813 | Improved |
| use lower bound E[log p] instead of log E[p] (ablation) | RandAR XL/16 · ImageNet-Val (K=20) | top-1 accuracy: 0.760 -> 0.813 | Improved |
| add latent noise augmentation to LlamaGen tokenizer (ablation) | RandAR L/16 · ImageNet-Val | top-1 accuracy: 0.769 -> 0.780 | Improved |
| add latent noise augmentation to LlamaGen tokenizer (ablation) | RandAR XL/16 · ImageNet-Val | top-1 accuracy: 0.812 -> 0.813 | No clear effect |
| use LlamaGen tokenizer instead of MaskGIT (ablation) | RandAR L/16 · ImageNet-A | top-1 accuracy: 0.101 -> 0.146 | Improved |
| score RandAR as any-order AR instead of masked diffusion (ablation) | RandAR-B · ImageNet-Val (K=20) | top-1 accuracy: 0.662 -> 0.708 | Improved |
| score as AR instead of masked diffusion at K=1 (ablation) | RandAR-B · ImageNet-Val | top-1 accuracy: 0.348 -> 0.586 | Improved |
| scale RandAR from L/16 to XL/16 | RandAR · ImageNet-Val | top-1 accuracy: 0.780 -> 0.813 | Improved |
| replace DiT classifier with RandAR | generative classifier XL/16 · ImageNet-C Gaussian | top-1 accuracy: 0.505 -> 0.652 | Improved |
| replace DiT with RandAR | generative classifier · Camelyon17 OOD (B/16) | accuracy: 0.586 -> 0.783 | Improved |
Closest papers by meaning
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.89)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.88)
- Stabilize the Latent Space for Image Autoregressive Modeling: A Unified Perspective (NeurIPS 2024 · similarity 0.88)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.88)
- Generative Classifiers Avoid Shortcut Solutions (ICLR 2025 · similarity 0.88)
- OmniGen-AR: AutoRegressive Any-to-Image Generation (NeurIPS 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.
