Does flux-dev or flux-fill support paired data training for LoRA?

[huxian0402]

I would like to train a LoRA model for object removal using paired object removal data based on flux-dev or flux-fill. The paired data includes the original image, the mask image of the object to be removed, and the target image (the image after the object has been removed). Does the current code support this? Do you have any suggestions? @lansing @rockerBOO @akx @tsukimiya @wkpark

[rockerBOO]

I developed a paired dataset system for preference optimization but it needs another pass to bring it up a good standard. The whole dataset system is a little confusing so might be an opportunity to consider improvements to allow for more flexibility. I was going to be making another pass on my preference optimization soonish so I might propose some improvements there that would allow paired or greater batches for doing associated images in training.

[hnsywangxin]

I have the same question. Did you solve it? @huxian0402