some bad results

[chiehwangs]

Hi,

When running splatting avatar on a new dataset, I encountered some strange results (as shown in the first fig, all results like this).

I believe my preprocessing should be fine, and the IMavatar fitting results also appear normal (fig.2 is image after RoubustVideoMatting, and fig.3 is output after imavatar preprocess pipeline. In my previous attempts, if there were errors in the intrinsic or extrinsic parameters, the 3DMM fitting on the right side of the last fig would not have been successful, but now, it seems fine). Where do you think the possible problem might lie?

[amerssun]

Even after retraining on the author's openly preprocessed data, I still encountered blurry effects (on the Bala dataset). While working with the Yufeng dataset, I successfully managed to replicate the face, but rotating the head led to abnormal facial voids.

[initialneil]

Enabling the scaling regularization will help to reduce artifacts on large viewing angle change.
Also, if the training images are front facing only, the algorithm will not generate well on large viewing angle changes.

[wengjincheng]

t

I am facing the same thing, I trained model using subject1 dataset, the evaluation result is bad. I don't know why it is too slow compare it with the data in paper.

{"psnr": 20.929635456630162, "ssim": 0.8153632400717055, "lpips": 0.2079089300973075, "n_gauss": 593713}