Announcements
Introducing Vella-1.0 Preview
October 21, 2024
We are introducing Vella-1.0 Preview, a new generative AI model for virtual try-on. Vella-1.0 Preview distinguishes itself by preserving original garment details, setting it apart from other generative AI models. This feature ensures that the intricate designs, patterns, and textures of clothing items are maintained with high fidelity to generate authentic virtual try-on images.
Vella-1.0 Preview is currently available for free on Vellaml.com during this beta period. The service is open to early-access users, and a wider release is planned in the coming months.
State-of-the-Art Virtual Try-On
Vella-1.0 Preview is our latest AI model for virtual try-on, delivering superior performance compared to our research preview model, IDM-VTON, which was released earlier this year. Our comprehensive evaluations have shown that Vella-1.0 Preview outperforms IDM-VTON and other current models in industry benchmarks, generating high-fidelity virtual clothing try-on images that more accurately represent the original garments.
We evaluated Vella-1.0 Preview on public datasets like VITON-HD and DressCode, using industry-standard metrics. Our assessment included low-level reconstruction, for which we employed LPIPS (Learned Perceptual Image Patch Similarity) and SSIM (Structural Similarity Index Measure). To evaluate high-level image similarity, we utilized the CLIP image similarity score. Additionally, we assessed image authenticity using the Fréchet Inception Distance (FID) score. This multi-faceted approach allowed us to thoroughly evaluate the model's performance across various aspects of image generation and similarity.
Our tests reveal that Vella-1.0 Preview has superior performance on both the VITON-HD and DressCode datasets. The model outperforms other concurrent methods across diverse metrics, showcasing its versatility and robustness.
Real-World Scenarios
To showcase Vella-1.0 Preview's capabilities in real-world applications, we've conducted tests in diverse and challenging scenarios. Using the In-the-Wild dataset, we compared Vella-1.0 Preview with other diffusion-based VTON (Virtual Try-On) methods. The results showed that Vella-1.0 Preview outperforms other methods across all metrics, even surpassing the performance of IDM-VTON.
To further test the versatility and robustness of Vella-1.0 Preview, we conducted extensive experiments across a wide range of real-world conditions. Our tests encompassed a diverse array of body types, including plus-size models, petite frames, and athletic builds. We explored various poses, primarily focusing on static positions such as standing and sitting, as well as subtle movements. While we also tested more dynamic poses like walking or jumping, our model currently excels in static and subtly dynamic scenarios.
We evaluated the model's performance against clean, studio-like backgrounds as well as complex outdoor settings and busy indoor environments. Our testing included models from various racial and ethnic backgrounds, with diverse skin tones, facial features, and hair types. We pushed the boundaries by testing full outfits combining tops and bottoms, layered clothing, and various fabric types and patterns.
The results of these experiments were impressive within the scope of our testing. Vella-1.0 Preview consistently generated natural-looking virtual try-on images across this range of conditions, particularly excelling in static poses and controlled environments. We observed accurate adaptation to various body shapes, with the model maintaining proper fit and style regardless of body shape or build. The realistic garment draping in static poses was particularly noteworthy, as was the seamless integration of virtual clothing with different backgrounds.
Conclusion
Vella-1.0 Preview is the advanced AI model powering the Vella service, marking a significant advancement in virtual try-on technology. It excels at preserving original garment details while adapting to diverse body types, clothing styles, and environments, outperforming existing methods across key benchmarks and real-world scenarios.
Vella leverages this model to offer practical applications across multiple industries. In e-commerce, it enhances the online shopping experience by allowing customers to virtually try on clothing—reducing return rates and increasing satisfaction—while also helping sellers create promotional images without the need for extensive photoshoots. In advertising, it enables diverse marketing campaigns with custom models and outfits, making ads more effective and inclusive across different demographics. In entertainment, it assists in designing costumes for films, TV shows, and games, streamlining the design process, and bringing creative ideas to life quickly.
By integrating the capabilities of Vella-1.0 Preview, the Vella service addresses real-world challenges and offers practical solutions, transforming how industries approach fashion and visual media and improving efficiency and user experience without unnecessary embellishments.
What’s Next
We plan to release the Vella-1.0 Preview model as open source for the community soon, just as we did with IDM-VTON. This will allow developers and researchers to build upon our work and drive further innovation in virtual try-on technology. To enhance efficiency, we're optimizing the model to double its inference speed while maintaining performance, making it lighter and faster. Based on continuous improvements and user feedback, we aim to launch the official version of Vella-1.0 in a few months.
We're excited to see how you will utilize Vella and encourage you to try it out and share your feedback with us.
Author
OMNIOUS.AI