Inference-Time Scaling for Diffusion Models beyond Scaling Denoising Steps (2501.09732v1)

Abstract: Generative models have made significant impacts across various domains, largely due to their ability to scale during training by increasing data, computational resources, and model size, a phenomenon characterized by the scaling laws. Recent research has begun to explore inference-time scaling behavior in LLMs, revealing how performance can further improve with additional computation during inference. Unlike LLMs, diffusion models inherently possess the flexibility to adjust inference-time computation via the number of denoising steps, although the performance gains typically flatten after a few dozen. In this work, we explore the inference-time scaling behavior of diffusion models beyond increasing denoising steps and investigate how the generation performance can further improve with increased computation. Specifically, we consider a search problem aimed at identifying better noises for the diffusion sampling process. We structure the design space along two axes: the verifiers used to provide feedback, and the algorithms used to find better noise candidates. Through extensive experiments on class-conditioned and text-conditioned image generation benchmarks, our findings reveal that increasing inference-time compute leads to substantial improvements in the quality of samples generated by diffusion models, and with the complicated nature of images, combinations of the components in the framework can be specifically chosen to conform with different application scenario.

• The paper introduces a novel inference-time scaling framework that improves diffusion model performance by optimizing noise inputs.
• It formulates noise optimization as a search problem, evaluating multiple verifiers and algorithms to tailor generation tasks.
• Extensive experiments across image synthesis tasks demonstrate that increased inference computation yields significantly better sample quality.

Inference-Time Scaling for Diffusion Models beyond Scaling Denoising Steps

The paper "Inference-Time Scaling for Diffusion Models beyond Scaling Denoising Steps" explores the concept of inference-time scaling in diffusion models, a topic that has garnered significant interest within the field of generative modeling. The work investigates how these models, specifically in image generation, can benefit from increased computational resources during inference, beyond merely increasing the number of denoising steps.

Key Contributions and Findings

1. Inference-Time Scaling Framework: The authors propose a novel framework for scaling diffusion models at inference time, beyond the conventional method of simply increasing denoising steps. They introduce a search-based strategy that involves identifying optimal noise inputs for the diffusion process, which has shown promise in improving sample quality significantly.
2. Search Problem Formulation: The research formulates the task as a search problem in which the primary objective is to find "better" noise inputs for the diffusion process. The design space is structured along two primary axes: (a) the verifiers used to provide feedback on generated samples, and (b) the search algorithms used to navigate the noise space.
3. Experimental Validation: Extensive experiments demonstrate the efficacy of the proposed methods across various benchmarks, including class-conditioned and text-conditioned image generation tasks. The results indicate that additional computational investment at inference can enhance the performance of diffusion models across different scenarios, underscoring the potential for tailored search strategies based on specific generation tasks.
4. Evaluation of Verifiers and Algorithms: The paper thoroughly examines the impact of different verifier-algorithm combinations. Four distinct verifiers and three search algorithms are evaluated, showcasing that no single configuration is universally optimal. Instead, each task and application may require distinct settings to achieve the best results.
5. Analysis of Verifier-Task Alignment: The paper explores how various verifiers align with specific generation tasks, highlighting their inherent biases. This analysis suggests that different verifiers may have distinct strengths, such as capturing text alignment or prioritizing visual quality, which affects their performance in searching for optimal noise inputs.

Implications and Speculative Insights

The implications of this work are both practical and theoretical. Practically, the framework can enhance the quality of generated samples in applications such as image synthesis, providing a flexible approach to allocating inference-time computational resources. This could be particularly useful in contexts where generating high-quality outputs is crucial, such as digital content creation and virtual environment design.

Theoretically, the research pushes forward the understanding of how additional computation at inference can influence generative model performance, offering an alternative view to the training-time scaling laws that have traditionally guided model development.

Future Directions

The paper opens avenues for future research to explore the development of more sophisticated verifiers that can generalize across tasks and the design of task-specific search strategies. Additionally, the concept of scaling inference-time computation might extend beyond diffusion models to other generative modeling paradigms, prompting further investigation into its broader applicability.

Overall, this work provides a comprehensive examination of inference-time scaling in diffusion models, establishing a foundation for future advancements in leveraging computation to enhance the quality and utility of generated outputs in diverse applications.