jiaxue-ai
commented
3 months ago It occurs because the pixel loss exclusively applies to the unmasked area, while the masked region employs perceptual loss and adversarial loss, emphasizing perceptual similarity.
To ensure consistency in low-frequency color, consider retraining the model with a loss function tailored for color consistency in the masked region.
In the example of lama, I found that the texture of masked area is reasonable but with chromatic aberration. What could be the reason for this?
