Research map / Image generation / Fast diffusion sampling
Your VAR Model is Secretly an Efficient and Explainable Generative Classifier
Yi-Chung Chen, David I. Inouye, Jing Gao
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 |
|---|---|---|---|
| replace diffusion classifier with VAR-based A-VARC+ | DiT-based diffusion classifier · ImageNet-100 | accuracy / GFLOPs: 90.30 -> 89.32; 415056.0 -> 4649.4 GFLOPs (89x less) | Improved |
| add likelihood smoothing (S=10) (ablation) | VARC · ImageNet-100 | top-1 accuracy: 83.30 -> 88.26 | Improved |
| add CCA finetuning (ablation) | VARC · ImageNet-100 | top-1 accuracy: 83.30 -> 88.68 | Improved |
| add CCA finetuning (ablation) | VARC · ImageNet-Sketch / ImageNet-A | top-1 accuracy: IN-Sketch 36.0 -> 34.8; IN-A 10.3 -> 9.9 | Hurt |
| add smoothing and CCA together (ablation) | VARC · ImageNet-100 | top-1 accuracy: 83.30 -> 89.72 | Improved |
| increase smoothing samples beyond 10-16 (ablation) | A-VARC · ImageNet-100 | top-1 accuracy: 88.26 -> 88.42 (S=32) | No clear effect |
| increase smoothing noise sigma to 1.0 (ablation) | A-VARC · ImageNet-100 | top-1 accuracy: 88.26 -> 39.54 | Hurt |
| use VAR classifier for robustness under distribution shift (vs diffusion classifier) | A-VARC+ · ImageNet-Sketch | top-1 accuracy: DC 53.7 vs A-VARC+ 34.0 | Hurt |
| use rectified-flow MeanFlow classifier instead of DiT diffusion classifier | diffusion classifier · ImageNet-100 | top-1 accuracy: 86.32 -> 50.30 | Hurt |
| increase VAR model size from d16 to d30 | VAR classifier · ImageNet-100 | top-1 accuracy: 83.30 -> 64.96 | Hurt |
| use generative classifiers for class-incremental learning (no rehearsal) | A-VARC+ · first 10 ImageNet classes | average accuracy: 41.2 -> 77.4 | Improved |
| apply CCA finetuning before token-wise mutual information explanation (ablation) | A-VARC+ · ImageNet-100 insertion/deletion AUC | insertion AUC: 0.845 -> 0.944 | Improved |
Closest papers by meaning
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.87)
- Generative Classifiers Avoid Shortcut Solutions (ICLR 2025 · similarity 0.87)
- Continuous Visual Autoregressive Generation via Score Maximization (ICML 2025 · similarity 0.87)
- Fundamental Limits of Visual Autoregressive Transformers: Universal Approximation Abilities (ICML 2025 · similarity 0.86)
- TTS-VAR: A Test-Time Scaling Framework for Visual Auto-Regressive Generation (NeurIPS 2025 · similarity 0.86)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.85)
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.
