Warped cloth used in Leffa Loss

[dh-thanh]

Thank for your wonderful work.

I also reproduce Leffa Loss in your paper. But the warp cloth get is weird. Does it correct?

In my opinion, It's imposible to get the correct warped cloth as in your paper. The reason is the value in attention score is very small (maximum value is arround 0.1) -> Flowfield is small -> allway grid sample in small region in reference image

(the red circle is small region I mentioned, the green circle is the expected region)

I'm trying to reproduce intergrate Leffa Loss to IDM-VTON as in Table 4

• What was your SD (SD 1.5 or SDXL)?
• Did you train both Unet as in your paper in Leffa architecture?
  
  Can you share the config when IDM-VTON

[animemory]

mark