Enhanced Realism in Virtual Try-On Tasks Using Diffusion Methods

Virtual try-on technology is revolutionizing online retail by enabling customers to visualize garments on their bodies before purchasing. Traditional methods, often based on Generative Adversarial Networks (GANs), face challenges such as misalignment and visual artifacts, especially in complex poses...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:2025 11th International Conference on Computing and Artificial Intelligence (ICCAI) s. 128 - 133
Hlavní autoři: Kiattithapanayong, Saris, Phoomvuthisarn, Suronapee
Médium: Konferenční příspěvek
Jazyk:angličtina
Vydáno: IEEE 28.03.2025
Témata:
Accuracy
Autoencoders
Benchmark testing
Computer architecture
ControlNet
Diffusion models
Diffusions
Kernel
Training
Vectors
Virtual Tryon
Visualization
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Virtual try-on technology is revolutionizing online retail by enabling customers to visualize garments on their bodies before purchasing. Traditional methods, often based on Generative Adversarial Networks (GANs), face challenges such as misalignment and visual artifacts, especially in complex poses. We present a virtual try-on framework leveraging diffusion models to enhance realism, accuracy, and garment detail preservation. Our approach integrates Vector Quantized Variational Autoencoders (VQ-VAEs) for precise feature matching within a diffusion U-Net architecture. By adopting image-based conditioning with the CLIP image encoder, our system utilizes visual features directly from clothing images for more faithful garment representations. Additionally, an Additional Feature Preserving Block (ControlNet) maintains intricate details like textures and logos, addressing fine-grained garment fidelity challenges. Quantitative evaluation demonstrates our system's superior performance, achieving the best LPIPS of 0.082. We also achieve a Fréchet Inception Distance (FID) of 7.782 and Kernel Inception Distance (KID) of 1.53, indicating enhanced image quality and feature alignment. Although the Structural Similarity Index Measure (SSIM) of \mathbf{0. 8 2 5} is slightly lower, it underscores the trade-off for improved realism and garment detail preservation. Our contributions set a new benchmark for accurate and realistic clothing visualization in virtual try-on systems.
DOI:10.1109/ICCAI66501.2025.00028