Research map / Image generation / Text-to-image diffusion
Sample Reward Soups: Query-efficient Multi-Reward Guidance for Text-to-Image Diffusion Models
Yinghua Yao, Yuangang Pan, Guoji Fu, Ivor Tsang
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 SRSoup K=20 instead of weighted-sum guidance | SRSoup · Aesthetic & Compressibility | reward queries per prompt: 15,000 -> 10,200 | Improved |
| use SRSoup K=20 instead of weighted-sum guidance | SRSoup · Aesthetic & Compressibility | inference time per prompt (min): 28.8 -> 21.1 | Improved |
| use SRSoup instead of weighted sum for a matched hypervolume | SRSoup · Aesthetic & Compressibility / Aesthetic & PickScore | reward query efficiency: 1.8x fewer reward queries | Improved |
| use SRSoup instead of weighted sum for a matched hypervolume | SRSoup · Aesthetic & HPSv2 | reward query efficiency: about 2.7x | Improved |
| add sample reward soup steps in the early denoising stage (K soup steps) (ablation) | SRSoup · Aesthetic & Compressibility | reward queries: reduces reward queries substantially (by up to 40%) | Improved |
| reduce K from 50 so weighted-sum search gradients enter earlier (ablation) | SRSoup · Aesthetic & Compressibility | Pareto front | Improved |
| increase queries per reward per step N (10, 20, 30) (ablation) | SRSoup · Aesthetic & Compressibility | Pareto front | Improved |
| use sample reward soup guidance instead of unconditional sampling for first K steps (ablation) | SRSoup · Aesthetic & Compressibility | Pareto front | Improved |
| replace fine-tuned model soups (DDPO, TDPO, AlignProp) with SRSoup | SRSoup · animal and HPD prompts, SD1.5 | Pareto front | Improved |
| replace AlignProp (soup) with SRSoup | SRSoup · Aesthetic & Compressibility, CLIP w1=0 | CLIP score: 0.1844 -> 0.2692 | Improved |
| replace TDPO (soup) with SRSoup | SRSoup · Aesthetic & HPSv2, CLIP w1=1 | CLIP score: 0.1128 -> 0.2324 | Improved |
| replace FT-WS, RiC, D3PO (soup) and Flow-GRPO soup with SRSoup | SRSoup · SD1.5 / SD3 multi-reward | Pareto front | Improved |
Closest papers by meaning
- Inference-Time Alignment of Diffusion Models with Direct Noise Optimization (ICML 2025 · similarity 0.92)
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 2025 · similarity 0.91)
- DiffExp: Efficient Exploration in Reward Fine-tuning for Text-to-Image Diffusion Models (AAAI 2025 · similarity 0.90)
- Diffusion Tree Sampling: Scalable inference‑time alignment of diffusion models (NeurIPS 2025 · similarity 0.90)
- Test-time Alignment of Diffusion Models without Reward Over-optimization (ICLR 2025 · similarity 0.90)
- A Dense Reward View on Aligning Text-to-Image Diffusion with Preference (ICML 2024 · similarity 0.89)
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.
