NEW BEST Virtual Try-On Models! Sm4ll VTON

Good morning everyone! We wanted to bring to attention a rather unknown update in the AI space for virtual try-on and virtual fashion. Today

Good morning everyone!

We wanted to bring to attention a rather unknown update in the AI space for virtual try-on and virtual fashion. Today we are taking a look at Sm4ll VTON's, a collection of VTON models created by Andrea Baioni & Alex Puliatti. See their original announcement video in the resources below!

AI fashion & consistency has been a problem for the better part of the last two years now. While there are tons of options to generate virtual try on images like Reactor, IPAdapter, IDM Vton, ACE ++, auto regressive models, etc. they all fall just a bit short. Whether it's incorrect product images, illegible text, requiring several generations for a usable image, or completely changing your source images, there really isn't a sure fire way to generate virtual try on images with reliable accuracy.

Sm4ll VTON's attempt to attack this problem by essentially compressing the entire masking and generation process into a collection (or family) of VTON models.

While existing methods try to encompass all image editing under single comprehensive models, Sm4ll VTON has a variety of models catered to individual use cases, whether it's upper body, lower body, shoes, eyewear, or full body outfits, Sm4ll VTON is capable of making intricate and accurate changes without effecting the original composition of your source images.

Currently, the models are not open-source, but Andrea and his team have mentioned discussions on providing open-weights for some of the models. You can access Sm4ll VTON via the Hugging Face Demo in the resources below!

Our $0.02

I think these models are fantastic, but I don't know if they are fully production ready yet as is. While these models are essentially the "end of the road" in terms of virtual try on as we understand it, I think there's more area to be explored for detailed, patterned, unique clothing designs.

Sm4ll VTON does a great job at streamlining and centralizing what we already knew into a clean interface and simple to understand process, but it doesn't bring anything new or solve any existing problems - it just more accurately completes what we have already been doing in virtual try on.

That's not to downplay the importance of the model(s), I think they are fantastic and certainly better than we have today, but even with these VTON models, there's still really no "production ready" system for consistent virtual try on generation over generation out of the box, but the gap is closing!

Resources:

Andrea's Sm4ll VTON Announcement: https://www.youtube.com/watch?v=5o1OjWV4gsk

Sm4ll VTON Technical Paper: https://sm4ll-vton.github.io/sm4llvton/

Sm4ll VTON Hugging Face Demo: https://huggingface.co/spaces/sm4ll-VTON/sm4ll-VTON-Demo