How does the identity weight affect the outcome?

[taki0112]

How does the identity weight affect the outcome?

For the current code, you gave a weight of 0.5
Did you try this when you gave it 1? What's the difference?

[taesungp]

It does not result in a big difference.

With higher identity loss, the image translation becomes more conservative, so it makes less changes. I tried using values like 0.1, 0.5, 1 and 10, and the result is not so different. I believe this is because the generator figures out essentially two different translation functions by detecting whether the input image is from the source dataset or the target dataset.