Subclass Feature Evaluation of Latent Diffusion Model
Che Bin Foo, Christopher Lauren Ann · 2023
Diffusion models have garnered success in unsupervised/semi-supervised learning. The synthetic images generated from this model can be used to augment test data for validation and verification. To trust these generated data for this use, it is important to develop methods to evaluate and measure these diffusion models and the generated images. In our research, we evaluated a large pre-trained text-conditioned latent diffusion model. The evaluation was done in both the latent space of the diffusion model and the pixel space of the generated images. The results show that the corresponding features generated are diverse, we can use the metrics provided from this research to measure this diversity in the latent and pixel spaces.