Bad Generated Images

[nazapip]

Hi,
Appreciate the great work and contribution.
I tested the ladi-vton model on a large number of images on VITON-HD dataset. Some of them which are not working properly i am sharing below

1. if the model image 'I' is wearing a full sleeves cloth and if i try to replace it with sleeveless or half sleeves, the portion of sleeve remains on the body [image 2].
2. It kind of tries to inpaint the mask portion exactly which doesn't look perfect in some images and the mask portion is visible at the bottom where the style is not in-shirt type
3. The occlusion part is not handled perfectly , the images with occlusion are not generate perfectly, in fact results are very distorted [image1]
4. And yes, the texture information is not preserved of the cloth image properly [image4]

What could be the reasons for these problems and does finetuning or training with images on a larger number of sleeve sleeveless combination would resolve this issue?