Usability of Pre-trained Diffusion Models in Generating Novel Datasets and Its Performance Evaluation

Md. Yearat Hossain, Md. Mahbub Hasan Rakib, Ifran Rahman Nijhum

Abstract: Even though sophisticated deep learning methods are getting better and better day by day, still they rely on a large number of datasets. But it is not always possible to acquire large datasets for all kinds of problems. Though diffusion models are now popular for their creative applications, it is already proven that they can generate better realistic-looking synthetic images compared to Generative Adversarial Networks (GAN). GANs are a popular option for image synthesis that helps the data sampling process for datasets that have low amounts of data or imbalanced data. In our work, we have experimented with a pre-trained text-to-image generation diffusion model for generating datasets for two different classes of problems. These problems are two common problems that can get benefitted from deep learning-based solutions but the lack of datasets hampers the process. We used the diffusion model to generate synthetic images and used those images as the training and validation data for the problems we tried to solve. Then we tested the models with manually collected real-world data and demonstrated the performance of such a method comparatively. From our experiments, we found that the diffusion model can generate realistic images and is up to 50 times faster in data generation compared to the manual human process. Also, in our testing, we found that the Convolutional Neural Networks trained with these synthetic data can achieve up to 80% and 89% accuracy scores.

External IDs:doi:10.1145/3594409.3594430