Support for exact clothes like ip-adapter-face model?

#14
by TerraNull - opened

Sorry if this is the wrong place to ask but is it possible to create an ip-adapter-clothes model like the ip-adapter-face model?

I've been trying to import specific clothes into people but while it does get there 70-80% with the general-ip-adapter on sdxl model, it misses important features that defines the clothes.

The ip-adapter-face works really well so I'm wondering if it's also possible to do this for clothes itself and have similar consistency as well.

I think it is possible https://github.com/tencent-ailab/IP-Adapter/pull/135#issuecomment-1803437109

Ya this is the problem, I've been able to get results like this before (which is like 70-80% there).
image.png
These images tend to be less consistent;
example1 : this image shifts colors and the logo changes dramatically.
Conditioning Image:
image.png
Output:
image.png

example2 (this image adds extra features such as a different collar):
Conditioning Image:
image.png
Output Image:
image.png

image.png

The reason I brought it up was because the face models perform extreamly well at least for me.
So I was thinking that if there was a seperate ip-adapter-clothes model it might perform that well, then again I don't know how this works.

yes, cloth is more difficult compared to face

yes, cloth is more difficult compared to face

Ya, I might just have to wait till something can do it. It's nothing critical. Just woul be so cool.
Anyways thanks for all the work you do.

TerraNull changed discussion status to closed

Thanks a lot.
In fact, cloth model is more important, it will have a huge impact on the apparel industry if it succeeds. we are planning to try it internally, but it is limited by the dataset (no large paired image dataset).

any update on this?

Sign up or log in to comment