Judging RGB Noise Test Method

[faruknane]

Hi @guangkaixu,

Thank you for the brilliant paper! Analyses are really good there. I was also training a diffusion segmentation model which I noticed it was shortcutting as well since It was doing the all job at the first backward step. Apart from that, I just wanted to ask your testing procedure when it comes to RGB noise with multiple steps. I'm thinking that in multi step approach, the image that is fed to the model starts converging from RGB image to Target Image (lets say we are predicting depth). While this is happening, the images in the intermediate steps starts becoming not very meaningful since they are mixture of RGB input and the depth prediction. I think this introduces ambiguity since the color shift in the image casts a cloud on the RGB image which makes it harder for model to understand the image. I'm wondering, do you also separately feed the original RGB image intact to the model with the current latent code $x_{t}$? If that's not the case, I think feeding it would be a better option. Please notice that DepthFM (@mgui7 @joh-schb) is feeding the original image at the side as well as the current latent code if I am not mistaken. What do you think about this?