Abstract
Virtual Try-On (VTON) has traditionally relied on an inpainting paradigm that utilizes binary masks to isolate garment areas. However, this approach frequently introduces artifacts, such as "fabric residues" or distorted limbs, due to the loss of underlying spatial and skin information. We introduce MFTON (Mask-Free virtual Try-ON), a novel framework that abandons the reliance on inpainting masks in favor of a mask-free denoising strategy. By incorporating a clothing texture-aware attention module and a geometric detail capture module, MFTON preserves person identity and intricate clothing textures with high fidelity. Experimental results on VITON-HD and DressCode datasets demonstrate that MFTON consistently outperforms existing state-of-the-art methods both quantitatively and qualitatively.
Introduction
The standard approach to Virtual Try-On involves creating a "clothing-agnostic" mask over a person's original attire and tasking a generative model with filling the masked region with a new garment. This "masking dilemma" presents significant challenges: simple masks may delete essential details like hands or hair, while complex parser-generated masks are prone to errors in challenging poses. Furthermore, masking removes critical context, forcing models to guess skin tones and spatial structures.
MFTON addresses these issues by shifting the paradigm from "cutting and filling" to "refining and blending." Instead of hiding parts of the image, the framework utilizes the original reference image as a denoising condition, allowing the model to distinguish between areas to keep and areas to replace.
Methodology
Mask-Free Denoising Strategy
MFTON utilizes the original Reference Person (RP) image as a conditioning input during the denoising process. By comparing a "Temporary Try-on Result"—a rough initial overlay—against the original person, the model learns to identify the boundaries between skin and fabric without the need for an explicit binary mask. This approach preserves fine details like hair overlapping with clothes and tattoos.
Clothing Texture-aware Attention
To capture high-frequency fabric patterns, MFTON introduces a texture-aware attention module. It employs an isotropic Laplacian filter to extract a texture map of the target garment, which is then injected into the diffusion model’s cross-attention layers. This ensures the model remains sensitive to specific weaves and patterns that global descriptors often miss.
Geometric Detail Capture and Consistency
To handle logos, text, and sharp boundaries, the framework incorporates a loss function based on the Sobel operator, penalizing gradient differences between generated and ground-truth images. Additionally, an appearance consistency inference method is used, starting the diffusion process from a posterior Gaussian noise rather than pure random noise. This reduces stochasticity and improves the stability of the final output.
Experimental Results
The model was evaluated on benchmark datasets including VITON-HD and DressCode. MFTON achieved a state-of-the-art FID of 9.382 and an SSIM of 0.902 on VITON-HD, indicating superior structural and distributional similarity compared to 30 other methods.
Qualitatively, MFTON excels in preserving difficult regions such as fingers and long hair. As shown in the comparison, competing models often struggle with complex text or produce blurred textures, whereas MFTON maintains sharp geometric boundaries and consistent fabric appearance.
Conclusion
MFTON demonstrates that providing more information—the original clothed person—is more effective than the traditional masking approach. By treating texture and geometry as distinct mathematical problems and reducing sampling randomness, the framework sets a new standard for high-fidelity virtual try-on. Future work will explore scaling these capabilities to more diverse "in-the-wild" scenarios and extreme body poses.
Summary updated