On the Noise Scheduling for Generating Plausible Designs with Diffusion Models

Jiajie Fan,Laure Vuaille,Thomas Bäck,Hao Wang

2023-11-19

Abstract:Deep Generative Models (DGMs) are widely used to create innovative designs across multiple industries, ranging from fashion to the automotive sector. In addition to generating images of high visual quality, the task of structural design generation imposes more stringent constrains on the semantic expression, e.g., no floating material or missing part, which we refer to as plausibility in this work. We delve into the impact of noise schedules of diffusion models on the plausibility of the outcome: there exists a range of noise levels at which the model's performance decides the result plausibility. Also, we propose two techniques to determine such a range for a given image set and devise a novel parametric noise schedule for better plausibility. We apply this noise schedule to the training and sampling of the well-known diffusion model EDM and compare it to its default noise schedule. Compared to EDM, our schedule significantly improves the rate of plausible designs from 83.4% to 93.5% and Fréchet Inception Distance (FID) from 7.84 to 4.87. Further applications of advanced image editing tools demonstrate the model's solid understanding of structure.

Computer Vision and Pattern Recognition,Artificial Intelligence,Computational Engineering, Finance, and Science

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to generate more reasonable structural designs by adjusting the noise schedule in the diffusion model. Specifically, the author focuses on how to ensure that the generated designs not only have high visual quality but also meet semantic rationality (i.e., no floating materials, missing parts, etc.) in the generation design tasks, especially in structural design generation. To achieve this goal, the author proposes the following points:

1. **Study the influence of noise level on the rationality of generation results**: The author finds that there is a specific range of noise levels, and the noise levels within this range have a decisive influence on the rationality of the generation results. Therefore, they propose a new noise scheduling method to optimize the training and sampling processes within this critical noise range.

2. **Propose two techniques to determine the critical noise range**: Through statistical analysis, the author proposes two techniques to determine the noise level range related to rationality. One is based on the Shapiro - Wilk test, and the other is based on the Kullback - Leibler divergence (KL divergence), which are respectively used to determine the end point (\(\sigma_{\text{end}}\)) and the starting point (\(\sigma_{\text{start}}\)) of the noise.

3. **Improve the noise schedule to enhance rationality**: The author modifies the method of noise scheduling in the training and sampling processes, so that more computing resources are concentrated in the noise range related to rationality. Experimental results show that this improvement significantly improves the rationality of the generated designs while maintaining good visual quality.

4. **Evaluate the performance of the generation model**: The author uses three evaluation metrics: Design Plausibility Score (DPS), Plausible Design Rate (PDR), and Fréchet Inception Distance (FID). Experimental results show that, compared with other advanced diffusion models, such as EDM, the Plausibility - oriented Diffusion Model (PoDM) proposed by the author performs better in generating reasonable designs, especially the PDR is increased from 83.4% to 93.5%, and the FID is also improved, from 7.84 to 4.87.

In conclusion, the core problem of this paper is to improve the rationality of the generated structural designs by optimizing the noise schedule in the diffusion model, thereby generating design schemes that are more in line with actual needs.