Why is the image treated as an additional condition?

[jingheya]

Thanks for sharing this interesting work!

In flow matching, the learned optimal transport function T maps a distribution x0 to another distribution x1. In your work, x0 represents the image distribution, and x1 is the depth distribution. It seems that, according to the principle of flow matching, we could directly map an image to a depth map. Why you still input the image into the UNet and treat it as an additional condition? Is this condition unnecessary?

Best,