search

miccunifi / ladi-vton

[ACM MM 2023] - LaDI-VTON: Latent Diffusion Textual-Inversion Enhanced Virtual Try-On
Other
412 stars 56 forks source link

Bad Result on custom image from DressCode Dataset #49

Open Shivanshmundra opened 8 months ago

Shivanshmundra commented 8 months ago

Hi Folks,

I tried inferencing on single image taken from dresscode with all preprocessed data from the original source data itself with minor tweaks. I am getting unexpected results from the custom inference.

Even when I am doing preprocessing myself the results are similar. I have attached input and output images as reference.(Pose map has 18 channels so couldn't visualize it properly here).

image image

Can anyone help me here?

Shivanshmundra commented 8 months ago

I am getting weird warped cloth in the process which might be causing this error

image
DevelMayCry-MrChen commented 8 months ago

Hi Folks,

I tried inferencing on single image taken from dresscode with all preprocessed data from the original source data itself with minor tweaks. I am getting unexpected results from the custom inference.

Even when I am doing preprocessing myself the results are similar. I have attached input and output images as reference.(Pose map has 18 channels so couldn't visualize it properly here).

image image Can anyone help me here?

i can help you, can you give me dresscode dataset?

Shivanshmundra commented 8 months ago

@DevelMayCry-MrChen thanks for helping out. I was able to run basic inference with some bug fixes. Still I think the model is not able to transfer model logos or text perfectly. image

Cloth image link and model image link. Let me know if you're able to process it well.

benemana commented 8 months ago

Hi @Shivanshmundra , it seems you were able to produce a pretty good result. Can you please share the code you used for inference? Thanks!

Shivanshmundra commented 7 months ago

@benemana you can check this issue for reference - https://github.com/miccunifi/ladi-vton/issues/25. I just followed preprocessing steps mentioned here, changed it to dresscode dataset though.