[ICML 2026 Spotlight] Lookahead Sample Reward Guidance for Test-Time Scaling of Diffusion Models

Authors: Yeongmin Kim, Donghyeok Shin, Byeonghu Na, Minsang Park, Richard Lee Kim, and Il-Chul Moon

ToDo: Add SDXL/FLUX backbones

Overview

Lookahead Sample Reward Guidance (LiDAR) is a test-time scaling method for diffusion model reward alignment. The performance increases with:

1. Higher lookahead accuracy (num_inference_steps $\uparrow$ in phase 1)
2. More lookahead samples (num_particles $\uparrow$ in phase 1)

Phase 1) Lookahead sampling and reward annotation

python lookahead_sampling.py --seed=100 --num_particles=50 --num_inference_steps=5 --save_individual_images=True

Phase 2) LiDAR sampling for target reward-tilted distribution

python LiDAR_sampling.py --use_rag --scale=12.5 --resample_t_end=200 --lookahead_path=100_50_5 --save_individual_images

Acknowledgements

This codebase builds upon and is inspired by:

• Diffusers: https://github.com/huggingface/diffusers
• ImageReward: https://github.com/THUDM/ImageReward
• FK-Diffusion-Steering: https://github.com/zacharyhorvitz/Fk-Diffusion-Steering
• DATE: https://github.com/aailab-kaist/DATE

Citation

@article{kim2026lookahead,
  title={Lookahead Sample Reward Guidance for Test-Time Scaling of Diffusion Models},
  author={Kim, Yeongmin and Shin, Donghyeok and Na, Byeonghu and Park, Minsang and Kim, Richard Lee and Moon, Il-Chul},
  journal={arXiv preprint arXiv:2602.03211},
  year={2026}
}