Research map / Image generation / Text-to-image diffusion
Diffusion Blend: Inference-Time Multi-Preference Alignment for Diffusion Models
Min Cheng, Fatemeh Doudi, Dileep Kalathil, Mohammad Ghavamzadeh, Panganamala Kumar
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 |
|---|---|---|---|
| blend backward diffusion of reward-specific fine-tuned models (DB-MPA) instead of averaging weights (Rewarded Soup) | Stable Diffusion v1.5 · Short-DrawBench w=0.5 | weighted reward: outperforms RS by 3.92x | Improved |
| sample LoRA adapters per denoising step (DB-MPA-LS) instead of evaluating all models (ablation) | DB-MPA · Short-DrawBench w=0.5 | weighted reward / time: 0.42 -> 0.39; 11.11 -> 5.64 sec/img | No clear effect |
| increase number of rewards from 2 to 4 (ablation) | DB-MPA-LS vs DB-MPA · Multi-reward | DB/LS gain ratio: 1.02, 1.15, 1.08 for 2, 3, 4 rewards | No clear effect |
| increase number of rewards from 2 to 4 (ablation) | Rewarded Soup · Multi-reward | DB/RS gain ratio: 3.3, 6.4, 8.2 | Hurt |
| blend pretrained and fine-tuned drift with factor lambda (DB-KLA) | Fine-tuned SD v1.5 (alpha=0.1) · Short-DrawBench | average reward | Improved |
| use lambda > 1 in DB-KLA | DB-KLA · Short-DrawBench | semantic accuracy | Improved |
| add conflicting JPEG compressibility reward | DB-MPA · Short-DrawBench | reward Pareto | No clear effect |
| scale to GenEval and SDXL | DB-MPA · GenEval 700 held-out prompts | Pareto front | Improved |
| use DB-MPA-LS sampling for KL control (ablation) | DB-KLA · KL alignment | feasibility | No clear effect |
Closest papers by meaning
- Diffusion Model Alignment Using Direct Preference Optimization (CVPR 2024 · similarity 0.92)
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 2025 · similarity 0.92)
- Aligning Diffusion Models by Optimizing Human Utility (NeurIPS 2024 · similarity 0.92)
- Ranking-based Preference Optimization for Diffusion Models from Implicit User Feedback (NeurIPS 2025 · similarity 0.91)
- DyMO: Training-Free Diffusion Model Alignment with Dynamic Multi-Objective Scheduling (CVPR 2025 · similarity 0.91)
- Training Diffusion Models with Reinforcement Learning (ICLR 2024 · similarity 0.91)
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.
