Open hangnima opened 2 months ago
Thanks for your interest in our work.
Due to the limited training resolution of FlashFace, which is only 768, the latent features corresponding to small faces are also relatively small. It is difficult to maintain good quality for fine facial structures like faces.
On the other hand, during the FlashFace training process, faces smaller than 1/7 of the image were removed.
If you want to achieve better facial results for small faces, please set the face bounding box and increase the fidelity-related hyperparameters. You can refer to our tutorial for guidance.
Thank you for the open source code for learning. I have a question to ask. I found that the consistency of the faces generated in close range (when the characters are close to the camera) is very high, but in long range (when the characters are far from the camera), the faces are prone to collapse