Research map / Reinforcement learning / RL from feedback
PromptLoop: Plug-and-Play Prompt Refinement via Latent Feedback for Diffusion Model Alignment
Suhyeon Lee, Jong Chul Ye
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 |
|---|---|---|---|
| train the MLLM policy with GRPO instead of using it untrained (ablation) | PromptLoop on SD1.5 · SD1.5 ImageReward | ImageReward: -0.2315 -> 0.4344 | Improved |
| use an untrained MLLM to refine prompts (ablation) | SD1.5 · SD1.5 ImageReward | ImageReward: 0.0816 -> -0.2315 | Hurt |
| add multiple prompt refinements (5 steps) within one diffusion trajectory (ablation) | PromptLoop on SD1.5 · SD1.5 ImageReward | ImageReward: 0.4344 -> 0.4912 | Improved |
| add training-time visual (latent) feedback (ablation) | PromptLoop on SD1.5 · SD1.5 ImageReward | ImageReward: 0.4912 -> 0.6320 | Improved |
| increase the number of prompt refinement steps at inference (ablation) | PromptLoop · SDXL and SD1.5 ImageReward | ImageReward | Improved |
| more refinement steps when trained without visual feedback (ablation) | PromptLoop · SDXL and SD1.5 ImageReward | ImageReward | No clear effect |
| add PromptLoop on top of Diffusion-DPO fine-tuned model | SDXL Diffusion-DPO · SDXL ImageReward | ImageReward: 0.9921 -> 1.2898 | Improved |
| add PromptLoop on top of NPNet | SDXL NPNet · SDXL ImageReward | ImageReward: 0.7357 -> 1.1213 | Improved |
| apply policy trained on SD1.5 to unseen SDXL | PromptLoop · SD1.5-trained, SDXL eval ImageReward | ImageReward: 0.7244 -> 1.0859 | Improved |
| replace PromptLoop by GPT-5 or Qwen2.5-VL-3B prompt refinement on SD1.5 | SD1.5 · SD1.5 ImageReward | ImageReward: 0.0816 -> -0.0950 (GPT-5), -0.1720 (Qwen2.5-VL-3B) | Hurt |
| PromptLoop vs RePrompt under composite reward | SDXL-turbo · SDXL-turbo composite reward | ImageReward: 0.7876 -> 0.8516 | Improved |
| use Qwen2.5-VL-3B prompt refinement without RL on SDXL-turbo | SDXL-turbo · SDXL-turbo composite reward | ImageReward: 0.7769 -> 0.6417 | Hurt |
Closest papers by meaning
- Aligning Diffusion Behaviors with Q-functions for Efficient Continuous Control (NeurIPS 2024 · similarity 0.89)
- Designing Reinforcement Learning for Diffusion Models: A Unified Path-Space View (NeurIPS 2026 · similarity 0.89)
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 2025 · similarity 0.88)
- Exploring Data Scaling Trends and Effects in Reinforcement Learning from Human Feedback (NeurIPS 2025 · similarity 0.88)
- Learning to Sample Effective and Diverse Prompts for Text-to-Image Generation (CVPR 2025 · similarity 0.88)
- PRDP: Proximal Reward Difference Prediction for Large-Scale Reward Finetuning of Diffusion Models (CVPR 2024 · 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.
