Research map / Image generation / Fast diffusion sampling
Lookahead Sample Reward Guidance for Test-Time Scaling of Diffusion Models
Yeongmin Kim, Donghyeok Shin, Byeonghu Na, Minsang Park, Richard Lee Kim, Il-chul Moon
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 |
|---|---|---|---|
| replace gradient guidance (DATE) with LiDAR (DPM-8, n=50) | SDXL · GenEval prompts | GenEval: 0.570 -> 0.585 | Improved |
| use a distilled one-step DMD model as lookahead sampler (ablation) | LiDAR on SDXL · GenEval prompts | GenEval: 0.585 -> 0.598 | Improved |
| increase lookahead samples n from 3 to 50 (DPM-5) (ablation) | LiDAR on SD v1.5 · GenEval prompts | GenEval: 0.449 -> 0.478 | Improved |
| add LiDAR on top of DPO fine-tuning | DPO-tuned SD v1.5 · GenEval prompts | GenEval: 0.446 -> 0.489 | Improved |
| use Universal Guidance (gradient) with ImageReward | SD v1.5 · GenEval prompts | GenEval: 0.426 -> 0.355 | Hurt |
| add negative guidance from low-reward lookahead samples (Safe-D, SR) (ablation) | SD v1.5 · ImageReward | ImageReward: 0.001 -> 0.014 | No clear effect |
| increase vanilla sampling steps from 100 to 250 | SD v1.5 vanilla · GenEval prompts | GenEval: 0.426 -> 0.430 | No clear effect |
| mix 10% CLIP into the IR reward annotation (ablation) | LiDAR · SD v1.5 GenEval prompts | GenEval: 0.478 -> 0.473 | Mixed |
| use CLIP only as reward instead of IR (ablation) | LiDAR · SD v1.5 GenEval prompts | GenEval: 0.478 -> 0.454 | Hurt |
| increase lookahead samples n on FLUX (ablation) | LiDAR on FLUX · GenEval | GenEval: 0.645 -> 0.668 | Improved |
| increase lookahead samples n on discrete diffusion (ablation) | LiDAR on UDLM (QM9) · ring count reward | novel ring count: 2.192 -> 4.128 | Improved |
| add LiDAR to SMC, BoN and SMC+BoN | SMC / BoN · SD v1.5 ImageReward, N=64 | ImageReward: 1.221 -> 1.338 (SMC) | Improved |
Closest papers by meaning
- Test-time Alignment of Diffusion Models without Reward Over-optimization (ICLR 2025 · similarity 0.89)
- Hierarchical Variational Policies for Reward-Guided Diffusion (NeurIPS 2026 · similarity 0.89)
- A General Framework for Inference-time Scaling and Steering of Diffusion Models (ICML 2025 · similarity 0.87)
- VGB for Masked Diffusion Model: Efficient Test-time Scaling for Reward Satisfaction and Sample Editing (NeurIPS 2026 · similarity 0.87)
- Test-time Scaling of Diffusions with Flow Maps (NeurIPS 2026 · similarity 0.87)
- Inference-Time Alignment of Diffusion Models with Direct Noise Optimization (ICML 2025 · 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.
