Quality of Synthetic Dataset

We ablate different variants of the synthetic training set: a prior model trained on gray-scale images, on low-quality renderings, in diverse environments, renders without hair and without makeup. Note the robustness of the finetuning towards weaker priors.

The first row show shows examples of the synthetic images for training the prior model. The second row shows the initialization before finetuning after warm-up. The third row contains novel views of the finetuned models. All results are generated from three inputs.

Prior without Hair: In-the-wild Results

We show additional results for the prior model without hair on in-the-wild captures. We show the initial renderings after warm-up ("before finetuning") and the finetuned results. The finetuning adds some hair but the quality is below the default model, which is trained on data with hair (see above for results of the full model).