Download link: https://drive.google.com/drive/folders/1eJrXvda9ZwA8NYOLtvr4N-iJ1u9wZ27J?usp=share_link

Notice:

① As explained by the authors of EG3D (please see the tri-plane issue), ffhq512-128.pkl and ffhqrebalanced512-128.pkl were achieved using buggy [XY, XZ, ZX] planes. As they suggested in the issue, we trained all of our EG3D-based models using fixed tri-plane [XY, XZ, ZY]. So be careful if you use set reload_modules=True in EG3D generation, please make sure to use the [XY, XZ, ZY] triplane when you reload modules.

② var3-64.pkl and var3-128.pkl are finetuned by fixing the camera parameters that are inputted into the generator as $c_g = c_{avg}$. Please use $c_{avg}$ as the conditional camera to use the two models.

var1-64.pkl : The model $E^{Ours}_{var1}$ in our paper, with neural rendering resolution of $64^2$. Trained on FFHQ+LPFF.

var1-128.pkl : The model $E^{Ours}_{var1}$ in our paper, with neural rendering resolution of $128^2$. Trained on FFHQ+LPFF.

var2-64.pkl : he model $E^{Ours}_{var2}$ in our paper, with neural rendering resolution of $64^2$. Trained on FFHQ+LPFF-rebal.

var2-128.pkl : The model $E^{Ours}_{var2}$ in our paper, with neural rendering resolution of $128^2$. Trained on FFHQ+LPFF-rebal.

var3-64.pkl : he model $E^{Ours}_{var3}$ in our paper, with neural rendering resolution of $64^2$. Trained on FFHQ+LPFF.

var3-128.pkl : The model $E^{Ours}_{var3}$ in our paper, with neural rendering resolution of $128^2$. Trained on FFHQ+LPFF.

To provide a fairer comparison, we also retrained EG3D using [XY, XZ, ZY] plane on the FFHQ dataset:

ffhq-fixed-triplane512-64.pkl : EG3D model trained with FFHQ dataset and has tri-plane fixed (using [XY, XZ, ZY]), with neural rendering resolution of $64^2$. Trained on FFHQ.

ffhq-fixed-triplane512-128.pkl : EG3D model trained with FFHQ dataset and has tri-plane fixed (using [XY, XZ, ZY]), with neural rendering resolution of $128^2$. Trained on FFHQ.

Download link: https://drive.google.com/drive/folders/1N4Tx5AAECueV3YEgDl0YpuRT7QcMWGk3?usp=sharing

FFHQ_LPFF.pkl : The model $S^{Ours}_{var1}$ in our paper. Trained on FFHQ+LPFF.

FFHQ_LPFF_rebalanced_maxsize89590.pkl : The model $S^{Ours}_{var2}$ in our paper. We mistakenly achieved this model using max_size = 89590. Notice that the dataset was shuffled before being clipped to size=89590, so the pose distribution are not affected much in this pretrained model. We additionally achieved a model using the entire dataset, please see below. Trained on FFHQ+LPFF-rebal (max_size = 89590).

FFHQ_LPFF_rebalanced.pkl (recommended): This model was achieved using the same training strategy and dataset as FFHQ_LPFF_rebalanced_maxsize89590.pkl, but without a max size limit. Trained on FFHQ+LPFF-rebal.