There are a couple of possible reasons, which is already reported in here: #13

Check DECA's results, which is available in data/Custom/data/$SUBJECT_ID/flame_init/renders. Some people do not have DECA's results and run fitting.

And try the fitting again.

Thanks for the prompt response. Indeed it was an issue with the DECA step. I had not downloaded the deca_model checkpoint, and the process had not raised any error.

I've now succesfully trained the avatar with your example dataset. May I ask if the results shown or the github project page are with only 5 epochs of training (which is the default in the code)?, I ask because I get pretty decent results but not as good as yours. Thanks

thanks for your confirmation. I added a line to ensure the availability of deca_model.tar: line.

I used only 5 epochs for all videos in the project page. Actually, you can download the avatars from here

Understood. Then for some reason the avatar I get is slightly worse. See the picture below and compare it with the one above. Some details (e.g. the shoes black areas) are missing. Any suggestions on what I should review?

How did you make the avatar? Did you do all the fitting process by yourself? Or used provided pre-processed fitting data and just made the avatar based on the provided ones?

For the avatar I did everything by myself except for the segmentation. That is, the only initial data I had was the video recording and the masks.

For the SMPL-X parameters of the target video (tvxq_mirotic) I used the data provided by you.

could you first check the quality of several dumped videos from the phone scan (not the target video)? you can compare the video with mine. for example, keypoints_whole_body.mp4 and smplx_optimized_smoothed.mp4

Yep okay that makes sense. I will see if I can resolve any discrepancies between my intermediate outputs and yours, and may open another issue if I can't solve it. Feel free to close this one, and thank you!