Research map / Image generation / Text-to-image diffusion
Self-Improving Diffusion Classifiers with Minority Preference Optimization
Hyunsoo Kim, Jungmyung Wi, Soobin Um, Donghyun Kim, Suhyun Kim
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 |
|---|---|---|---|
| apply MiPO fine-tuning | SD 1.5 diffusion classifier · CIFAR10-C | accuracy: 57.72 -> 59.76 | Improved |
| apply MiPO fine-tuning | SD 1.5 diffusion classifier · SUN09 | accuracy: 64.69 -> 68.25 | Improved |
| apply MiPO fine-tuning | SD 2.0 diffusion classifier · ImageNet-Tiny | accuracy: 51.45 -> 55.25 | Improved |
| apply MiPO fine-tuning | SD 2.0 diffusion classifier · CIFAR10 | accuracy: 88.58 -> 86.98 | Hurt |
| apply MiPO fine-tuning | SD 1.5 / SD 2.0 diffusion classifier · LabelME and VOC2007 | accuracy: 62.90 -> 55.28 (LabelME SD1.5); 81.84 -> 76.81 (VOC SD1.5) | Hurt |
| apply MiPO fine-tuning | SD 1.5 · CIFAR10 minority group | accuracy: 73.07 -> 80.58 | Improved |
| add KL regularization (ablation) | MiPO · CIFAR10 minority group | accuracy: 78.58 -> 80.58 | Improved |
| add KL regularization (ablation) | MiPO · CIFAR10 majority group | accuracy: 82.08 -> 82.68 | Improved |
| update early timesteps (0-20) instead of full (0-50) (ablation) | MiPO · CIFAR10 minority group | accuracy: 80.58 -> 81.68 | Improved |
| update late timesteps (30-50) instead of early (ablation) | MiPO · CIFAR10 minority group | accuracy: 81.68 -> 73.77 | Hurt |
| MiPO LoRA vs DDIM for generation | SD 1.5 · MS-COCO generation | Minority Score: 0.2132 -> 0.2289 | Improved |
| MiPO vs DDIM for generation quality | SD 1.5 · MS-COCO generation | HPSv3: 5.7581 -> 4.4090 | Hurt |
Closest papers by meaning
- Don't Play Favorites: Minority Guidance for Diffusion Models (ICLR 2024 · similarity 0.88)
- Elucidating the design space of classifier-guided diffusion generation (ICLR 2024 · similarity 0.88)
- Not All Diffusion Model Activations Have Been Evaluated as Discriminative Features (NeurIPS 2024 · similarity 0.87)
- Diffusion-NPO: Negative Preference Optimization for Better Preference Aligned Generation of Diffusion Models (ICLR 2025 · similarity 0.87)
- Diffusion Classifiers Understand Compositionality, but Conditions Apply (NeurIPS 2025 · similarity 0.87)
- Noise Matters: Optimizing Matching Noise for Diffusion Classifiers (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.
