GeneFacePlusPlus
GeneFacePlusPlus copied to clipboard
yerfor
The result video will flicker when human's eyes close.
https://github.com/yerfor/GeneFacePlusPlus/assets/6760332/25ae30e9-59b9-4758-90ff-624a5bf99226
Very great project! I test my video. As shown in the video above, when I close my eyes, the video will flicker, which is referred to in the following figure. And the problem happened at the 3s,7s and 11s. I wonder what caused the problem.
Well, what about this? I can't find the reason, plz help me. I used 23s video for training and the drv_aud is 6s long. As you can see, the result is so strange~
res.mp4
23s length may not be enough. Try a longer one. I used a 3-min video to train my own model
23s length may not enough yes, but before this, I trained with 6s video and got good result. And what I want to know is that the training material MUST be consecutive? I mean, I formed 23s training video by joining two parts. Is this so crucial?
23s length may not enough yes, but before this, I trained with 6s video and got good result. And what I want to know is that the training material MUST be consecutive? I mean, I formed 23s training video by joining two parts. Is this so crucial?
Yes, consecutive video frames are very important.
same problem, the reason might be overfitting or lpips_loss. Because the problem do not occur when I test my low-step result. In my case,I turned the max training step=60000,lpips_loss_update=50000,and modified the ambient_loss_target. Sometimes the problem can be solved,but there is no universal solution. I just watch my tensorboard and change the config file when the validation result is strange.
same problem, the reason might be overfitting or lpips_loss. Because the problem do not occur when I test my low-step result. In my case,I turned the max training step=60000,lpips_loss_update=50000,and modified the ambient_loss_target. Sometimes the problem can be solved,but there is no universal solution. I just watch my tensorboard and change the config file when the validation result is strange.
Many Thanks! I will have a try.
Meanwhile, the NeRF implemented of the work belongs to modulation-based method. Does the percentage of frame of which human closing eye affect the high-dimension slicing? I am experimenting with it and will update result later.
hey you trained model using default config values ? In my case there is no blinking at all in the resulting video, can you please help
