Number of Training Identities

We ablate fine-tuning results for prior models trained on a lower number of synthetic identities. For each identity, we sample 13 expressions and 30 views per expression. For example, the prior model for the column "15 Subjects" is trained on 15 subjects x 13 expressions x 30 views, i.e., 5,850 training images. The videos show the RGB renderings and predicted normals.