Results not as good as expected

[cardosofelipe]

First things first: thank you for the great work, your project is one of the best open-sourced virtual-try-on via AI implementations I have seen so far. Despite the results I'm getting, I am still very impressed by the amount of work you put on it and the amount of material released.
Now, back on the issue.
In the last few days I am trying to test your model against some custom data. Initially I struggled to avoid some "bad" pictures for the pre-processing scripts, however I managed to obtain some model/product pairs and to build a small dataset to test the capabilities of such model. Some results are a bit weird, for instance,
this product:

on this model:

gives me this result:

The example above is one of the best results I got so far and it has several problems, but the one I would like to address with this issue is the texture. It seems that almost always I get a knitwear-like texture, even though the input piece of cloth is not one. These are some additional results:

Basically, not matter what kind of cloth I try, I always get a wrong texture, very similar to the ones shown above.

Technical Details:

• Image Format: JPG
• Image Dimensions: Width = 768, Height = 1024

Questions:

• Could the issue stem from the images I am using?
• Is there a potential bias in the model's weights?
• My inclination is to retrain the model with new data. Has this approach been attempted in response to similar issues?

I would greatly appreciate any suggestions or guidance on how to address this texture representation issue. Thank you for your time and assistance.

[ChiehYunChen]

Hi @cardosofelipe,

Thank you for sharing your experience.

The issue might be the checkerboard effect because our architecture contains "transposed convolution.”
If you have other cool findings, we appreciate knowing them!

[StartHua]

Me too, have you solved it?