OutfitAnyone Review
OutfitAnyone by HumanAIGC is a virtual try-on demo that emerged in December 2023. It allows users to simulate any clothing on an AI-generated model, achieving some very impressive results. This review will explore the demo's strengths and weaknesses, providing insights into its current performance and capabilities.
Written by Aya Bochman | February 26, 2024
Overview
OutfitAnyone is a new AI image model from Alibaba Group research team - Institute for Intelligent Computing. It allows you to instantly try on any outfit on a model for tops, bottoms, or full-body garments. You can even mix and match tops and bottoms, and it all takes less than a minute. You can freely upload any garment you want, but to prevent malicious use of personal photos, the models are from their own AI-generated selection, which, as of now, is pretty limited and not too diverse.
As stated on their GitHub page, this project was created for research purposes only, and no commercial benefits are to be derived from it. Most models and clothing images are taken from the internet, and they have trained their model on public datasets such as VITON and DressCode, which are estimated to consist of about 4-10 million training samples. This is a respectable amount of data, hence we are able to see high-quality results.
So, how does it work behind the scenes? Without getting too technical here, OutfitAnyone's solution utilizes a multi-stage diffusion model (You can read more about it on their GitHub page). Their demo is publicly available on HuggingFace, and you can experiment with it for free.
In this review, we will experiment with several try-on use cases, including:
A fashion model wearing a garment
A fashion model with different skin color
A simple garment-only image
A complicated garment
A complicated background for the model or garment
A Fashion Model Wearing a Garment
We wanted to start with the best, and this result impressed us the most out of the experiment. We chose a pretty complicated but interesting tank top worn by a professional fashion model, without selecting a lower garment.
The result we got here is almost perfect, except for the distorted background, which was taken from the garment image portraying some sort of city landscape. It would have been better to consider the try-on model's blank background instead, to preserve our desired result and avoid unnecessary painting attempts.
Another thing to note is that the output model's skin became a bit tanner (we will return to this in the next experiment), as if it were trying to resemble our input model. The lower garment part was left empty, and it did a pretty good job of inventing pants similar to what the input model is wearing. If we focus on the top part, the result is great and accurate.
A Fashion Model with Different Skin Color
The next model we chose is not only from another ethnicity, but also has a different body shape than OutfitAnyone’s AI models. The result here was a bit surprising.
First of all, the OutfitAnyone model’s body changed completely, meaning we are losing the purpose of the "try-on" here. It looks like it took on the body shape of the input model - not the result we were looking for. Secondly, you can strongly see here the skin color change that seems to have taken effect only on the model's body, with the face left out. Interesting.
A Simple Garment-only Image
Here are a few examples of generating an image from only the garment, which works pretty well, especially when the background is plain.
The quality of the result is not as good, with a bit of data leakage, but it's impressive that the model was able to take the garment off the hanger. In the next attempt, we challenged it with a sequin shirt and no background, and you can see the result is flawless and accurately imitates the pattern, light, and shadows, so that's a success!
A Complicated Garment
Take a look at the dress we chose here to challenge the model a bit more.
The result is as detailed as possible and looks very nice at first glance. However, let's consider it an edge case, as it is really challenging to take into account so many details, especially when the input image was not exactly 100% front-facing, making it hard for the model to imagine all the details. Also, once again, we can observe the skin color change.
A Complicated Background for the Model or Garment
By now we’ve realized that OutfitAnyone’s model takes the full input image of a garment/model, including its background. So, we conducted one last experiment to see how well it handles the background and, of course, the garment.
The garment looks good here, very accurate with the details of the pattern and color, and the background is of very high quality but with some hallucinated details, which is somewhat expected.
Conclusion
Overall, OutfitAnyone is doing an amazing job, currently one of the best in this domain.
It arrived right on time during the AI boom era to showcase a less talked about use-case of AI: virtual fashion models, virtual fittings, and the huge potential of AI in the fashion sector.
Imagine a day when brand owners won't have to conduct fashion photoshoots with their clothes, and this problem could be solved by hiring remote models, or rather just generating an AI model with the perfect, custom-made prompt that fits the brand's image.