Closed AustinXY closed 2 years ago
I have fine-tuned an AFHQ model from the FFHQ model. Since the NeRF network is frozen, the 3D shape is learned from face dataset. As such, the performance is limited.
When I am free, I will train a model on AFHQ from scratch. I think it will not be more difficult compared to human faces.
As for compcar, the hyperparameters may need to be re-adjusted. I currently have no plans to do this experiment.
Do you have quantitative evaluations of the model on some other datasets (compcar or afhq)? If so, how do they compare to StyleNeRF results?