Kaede Shiohara comments

Results 16 comments of


                                            Kaede Shiohara

About "tokenized_identity_first" and "tokenized_identity_last"

Sorry for the late reply. We follow the multi-token ("f"irst name and "l"ast one) strategy proposed in DreamIdentity and CelebBasis.

Save SelfBlendedImages

This kind of incorrect cropping tends to occur when there is something wrong with the face detector (ie, retinaface). You can start by visualizing detected face bounding boxes on face...

trainset

Hi, You can generate SBI by the training code we have released in this repository.

The AUC on the DFD dataset is not as high as stated in the paper.

Hi, We used the "raw" videos (not "c23") for the evaluation. Proprocessing for DFD is the same as the other datasets.

which graphics card should be used for training the SBI model

We used a single NVIDIA A100 (40GB).

ModuleNotFoundError: No module named 'retinaface.pre_trained_models

You should install retinaface. We provide dockerfile to prepare for environment. Please refer to `README.md`.