Research map / Image generation / Fast diffusion sampling
Censored Sampling of Diffusion Models Using 3 Minutes of Human Feedback
TaeHo Yoon, Kibeom Myoung, Keon Lee, Jaewoong Cho, Albert No, Ernest K. Ryu
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 |
|---|---|---|---|
| use reward ensemble plus universal guidance instead of single reward model (ablation) | censoring pipeline · MNIST crossed 7 | proportion of malign images (%): 1.30 -> 0.42 | Improved |
| use ensemble instead of union reward model (ablation) | censoring pipeline · MNIST crossed 7 | proportion of malign images (%): 0.98 -> 0.60 | Improved |
| add universal guidance (backward + recurrence) to ensemble (ablation) | censoring pipeline · MNIST crossed 7 | proportion of malign images (%): 0.60 -> 0.42 | Improved |
| use ensemble instead of single reward model (ablation) | censoring pipeline · LSUN church watermark | proportion of malign images (%): 3.02 -> 1.36 | Improved |
| add recurrence to ensemble (ablation) | censoring pipeline · LSUN church watermark | proportion of malign images (%): 1.36 -> 0.76 | Improved |
| single models trained on 30 malign + 30 benign vs union models (ablation) | reward model · LSUN church watermark | proportion of malign images (%): 3.84 -> 3.02 | Improved |
| use imitation learning (3 rounds) instead of training at once on 30 malign (ablation) | reward model · ImageNet tench faces | proportion of malign images (%): 6.8 -> 2.2 | Improved |
| use imitation learning (2 rounds) instead of training at once on 20 malign (ablation) | reward model · ImageNet tench faces | proportion of malign images (%): 10.7 -> 7.5 | Improved |
| add universal guidance to imitation learning round 3 (ablation) | censoring pipeline · ImageNet tench faces | proportion of malign images (%): 2.2 -> 1.0 | Improved |
| use guidance-based censoring instead of rejection sampling (threshold 0.8) (ablation) | censoring pipeline · ImageNet tench faces | proportion of malign images (%): 29.8 -> 2.2 | Improved |
| use ensemble + universal guidance instead of ensemble only (ablation) | censoring pipeline · Stable Diffusion embedded text | proportion of malign images (%): 3.86 -> 1.24 | Improved |
| use ensemble instead of single reward model (ablation) | censoring pipeline · Stable Diffusion embedded text | proportion of malign images (%): 7.10 -> 3.86 | Improved |
Closest papers by meaning
- CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation (NeurIPS 2025 · similarity 0.89)
- Positive-Unlabeled Diffusion Models for Preventing Sensitive Data Generation (ICLR 2025 · similarity 0.88)
- Reward-Instruct: A Reward-Centric Approach to Fast Photo-Realistic Image Generation (NeurIPS 2025 · similarity 0.87)
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 2025 · similarity 0.87)
- Training-Free Diffusion Model Alignment with Sampling Demons (ICLR 2025 · similarity 0.86)
- CADS: Unleashing the Diversity of Diffusion Models through Condition-Annealed Sampling (ICLR 2024 · 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.
