Research map / Reinforcement learning / RL from feedback
Diffusion Fine-Tuning via Reparameterized Policy Gradient of the Soft Q-Function
Hyeongyu Kang, Jaewoo Lee, Woocheol Shin, Kiyoung Om, Jinkyoo Park
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 consistency model as x0 predictor (ablation) | SQDF · Aesthetic reward | aesthetic score: 7.10 -> 7.87 | Improved |
| add replay buffer (ablation) | SQDF · Aesthetic reward | DreamSim diversity: 0.56 -> 0.58 | Improved |
| add replay buffer (ablation) | SQDF · Aesthetic reward | aesthetic score: 8.06 -> 7.87 | Hurt |
| add consistency model (ablation) | SQDF · Aesthetic reward | DreamSim diversity: 0.62 -> 0.58 | Hurt |
| use discount factor gamma<1 instead of gamma=1 (ablation) | SQDF · Aesthetic reward | alignment and diversity | Improved |
| use gamma<1 instead of gamma=1 (ablation) | SQDF · Aesthetic reward | aesthetic score | Hurt |
| use 2-step DDIM instead of Tweedie one-step x0 prediction (ablation) | SQDF · Aesthetic reward | optimization speed | Improved |
| use 4-step DDIM instead of 2-step (ablation) | SQDF · Aesthetic reward | training stability | Hurt |
| use SQDF-UCB instead of SEIKO-UCB | SQDF · Online black-box optimization | aesthetic score: 7.49 -> 7.87 | Improved |
| use SQDF-Bootstrap instead of SEIKO-Bootstrap | SQDF · Online black-box optimization | DreamSim diversity: 0.24 -> 0.53 | Improved |
| use SQDF instead of KL-augmented DDPO/DRaFT | SQDF · Text-to-image finetuning | reward vs alignment/diversity trade-off | Improved |
| use SQDF instead of ReFL/DRaFT | SQDF · Text-to-image finetuning | alignment and diversity | Improved |
Closest papers by meaning
- Inference-Time Alignment of Diffusion Models with Direct Noise Optimization (ICML 2025 · similarity 0.89)
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 2025 · similarity 0.89)
- EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.89)
- Test-time Alignment of Diffusion Models without Reward Over-optimization (ICLR 2025 · similarity 0.89)
- Training-Free Diffusion Model Alignment with Sampling Demons (ICLR 2025 · similarity 0.89)
- AccuQuant: Simulating Multiple Denoising Steps for Quantizing Diffusion Models (NeurIPS 2025 · similarity 0.88)
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.
