# Inference-Time Scaling for Diffusion Models beyond Scaling Denoising Steps

**Authors:** Nanye Ma, Shangyuan Tong, Haolin Jia, Hexiang Hu, Yu-Chuan Su, Mingda Zhang, Xuan Yang, Yandong Li, Tommi Jaakkola, Xuhui Jia, Saining Xie  
**arXiv:** [2501.09732](https://arxiv.org/abs/2501.09732)  
**Date:** 2025-01-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2501.09732v1/

## Key points

- Researchers studied inference-time scaling behavior of diffusion models to enhance generation performance
- Diffusion models allow adjusting computation through denoising steps, with performance gains plateauing after a threshold
- Introduced a search problem to identify better noises for diffusion sampling process
- Structured design space along two axes: verifiers used for feedback and algorithms for finding improved noise candidates
- Evaluation conducted on DrawBench and T2I-CompBench datasets for text-to-image model performance
- FLUX.1-dev model used as backbone, employing supervised verifiers like Aesthetic Score Predictor, CLIPScore, and ImageReward
- Increasing inference-time compute enhances sample quality in diffusion models, especially in complex image generation tasks
- Self-supervised verifiers less effective in text-to-image settings due to focus on visual quality over textual information
- Metrics from DrawBench and LLM used for comprehensive evaluation
- Study highlights how increased computation during inference can improve sample quality in diffusion models for text-to-image tasks

## Summary

, , , , 

In this study, the researchers delve into the inference-time scaling behavior of diffusion models, particularly focusing on how increased computation can enhance generation performance. Unlike Large Language Models (LLMs), diffusion models offer the flexibility to adjust inference-time computation through denoising steps, with performance gains plateauing after a certain threshold. To explore this further, the researchers introduce a search problem aimed at identifying better noises for the diffusion sampling process. They structure the design space along two axes: the verifiers used to provide feedback and the algorithms employed to find improved noise candidates. The evaluation is conducted on two datasets: DrawBench and T2I-CompBench, which assess text-to-image models' ability to handle complex prompts and generate high-quality images. The FLUX.1-dev model serves as the backbone for this study, representing state-of-the-art text-conditioned diffusion models. Various supervised verifiers are utilized to evaluate different aspects of generated images, including Aesthetic Score Predictor, CLIPScore, and ImageReward. Additionally, a Verifier Ensemble is created by combining these verifiers to expand evaluation capacity. The researchers find that increasing inference-time compute significantly enhances sample quality in diffusion models, especially in complex image generation tasks. Self-supervised verifiers are found to be less effective in text-to-image settings due to their focus on visual quality over textual information. Metrics from DrawBench are used alongside an LLM as a neutral evaluator for comprehensive evaluation. Overall, this study sheds light on how increased computation during inference can lead to substantial improvements in sample quality in diffusion models for text-to-image generation tasks. By leveraging a combination of verifiers and metrics tailored to specific evaluation needs, the researchers provide valuable insights into optimizing model performance in complex image generation scenarios.