Page 1 of 1

Concentrate training on only "bad" Images

Posted: Mon Nov 25, 2019 1:30 pm
by nexus7
Hello, would it be a option for the future to have the faces preview shown like now but if you mark certain pictures that do not do well, the network would concentrate on those in future calculations?

Re: Concentrate training on only "bad" Images

Posted: Mon Nov 25, 2019 3:56 pm
by deephomage
I'm not sure why you would want the network to train on "bad" images. If you're seeing poor quality in certain images in the previews, you should manually align to check the alignments for those images. Poor quality face recreations may also point to a lack of image quality or not enough diversity of pose, expressions, and lighting in your training set. Face profiles and extreme face angles need manual aligning.

Re: Concentrate training on only "bad" Images

Posted: Mon Nov 25, 2019 6:08 pm
by nexus7
Ok, thank you, it seems i misunderstood.
I have a question about alignments files. When i have trained a model and want to use it to make a video i select the alignments file for the video of the face in the video. So i point to the alignments file in the folder from the extraction. The error "no alignments found for xxxpicture skip, and then the video is unedited, looks like the source video.
On the other hand if i do not point to the file, leave blank, to default location, i get the error

"WARNING No Alignments file found. Extracting on the fly.
NB: This will use the inferior cv2-dnn for extraction and landmarks. It is recommended to perfom Extract first for superior results"

but the video gets created with the faceswap.
How can i use the alignments file for a better result?