Open Syndulla opened 11 months ago
Hi, I think you can try to adjust this hyper-parameter here. The tgt_scale is used to deal with large occlusion by reducing the L2 loss between the swapped face and the target face, which might harm the identity similarity in some cases. You can try to set the tgt_scale to lower value.
Thank you - have tried adjusting this parameter with not such big an impact on results. I will however give it another go and documents the results better. Will keep this issue open for now in case I come back with another related point on the parameter.
Cheers!
Hi Chaps,
After some poking around I have got this running. However, the quality of my results seem to be dramatically worse than what is demonstrated in your paper. Wondering if you have any thoughts on where I might be going wrong. Examples of source, target and result images below:
Any ideas or thoughts you might have would be most appreciated.