Page 1 of 1

Questions

Posted: Tue Jul 21, 2020 10:16 am
by lorenzosale
Hello,

I managed to create a good deepfake, but I still have 2 questions please:

Once my video is converted, I notice that there is a delimitation (more or less perceptible depending on the image) in the shape of a square. I'm not sure what to adjust to mitigate this (see image below) can you help me please?



I used the "extended" mask. Maybe I have to change the mask before the extraction?

My second question is:

I find that my deepfake is really blurry, yet my dataset is of good quality, and I stopped training at +150,000 iterations

Do I still have to continue the training? Or does it have nothing to do with training but adjustments to be made?

Training :

Trainer : Original
Batch Size : 16 (optimizer saving : on)
It : 160 000
Loss Function : ssim

Convert :
Color Adjustment : Match Hist
Mask Type : Extended
Scaling : None

Thank you very much once again for your help :)

Re: Questions

Posted: Wed Jul 22, 2020 9:12 am
by torzdf
lorenzosale wrote:
Tue Jul 21, 2020 10:16 am

Once my video is converted, I notice that there is a delimitation (more or less perceptible depending on the image) in the shape of a square. I'm not sure what to adjust to mitigate this (see image below) can you help me please?


Your best bet is to use the preview tool, and adjust the box blend settings/mask blend settings until you have a result you are happy with.
I find that my deepfake is really blurry, yet my dataset is of good quality, and I stopped training at +150,000 iterations

Do I still have to continue the training? Or does it have nothing to do with training but adjustments to be made?
It can be numerous reasons, but this is in training, yes. It can be down to quality/variety of data, choice of model used and the model having not been trained for long enough.

Re: Questions

Posted: Wed Jul 22, 2020 3:15 pm
by lorenzosale
Thank you very much for your answers, I will try this