On the Efficiency of Diffusion Models in Generating Plausible Designs
Abstract: Diffusion-based generative models have huge potential in creating novel structural images in generative design where the user heavily values the design plausibility, e.g, no floating material or missing part. However, such models often require many denoising steps to achieve satisfactory plausibility, resulting in high computation costs; when using much fewer steps, we can not ensure plausibility. This paper addresses this trade-off and proposes an efficient training and inference method that can achieve the same or better plausibility than existing models while reducing the sampling time. We determine the noise schedule based on the evolution of pixel-value distributions in the forward diffusion process. Compared to previous models, e.g., DDPM and EDM, our method concentrates the noise schedule at a range of noise levels that highly influence the structural modeling and hereby achieves high efficiency in inference without compromising the visual quality or design plausibility. We apply this noise schedule to the EDM method on two structural data sets, BIKED and Seeing3DChairs. On BIKED images, for instance, our noise schedule significantly improves the quality of generated designs: the rate of plausible designs from 83.4% to 93.5%; FID from 7.84 to 4.87, compared to EDM.
Submission Length: Regular submission (no more than 12 pages of main content)
Previous TMLR Submission Url: https://openreview.net/forum?id=qWohURtuHJ&nesting=2&sort=date-desc
Changes Since Last Submission: Carefully revise the paper in accordance with the author guidelines and TMLR template and move the first figure to after the beginning of the first section.
Assigned Action Editor: ~Jun_Zhu2
Submission Number: 4349
Loading