FitDiT: Advancing the Authentic Garment Details for High-fidelity Virtual Try-on

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Jiang, Boyuan, Hu, Xiaobin, Luo, Donghao, He, Qingdong, Xu, Chengming, Peng, Jinlong, Zhang, Jiangning, Wang, Chengjie, Wu, Yunsheng, Fu, Yanwei
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913583172943872
author Jiang, Boyuan
Hu, Xiaobin
Luo, Donghao
He, Qingdong
Xu, Chengming
Peng, Jinlong
Zhang, Jiangning
Wang, Chengjie
Wu, Yunsheng
Fu, Yanwei
author_facet Jiang, Boyuan
Hu, Xiaobin
Luo, Donghao
He, Qingdong
Xu, Chengming
Peng, Jinlong
Zhang, Jiangning
Wang, Chengjie
Wu, Yunsheng
Fu, Yanwei
contents Although image-based virtual try-on has made considerable progress, emerging approaches still encounter challenges in producing high-fidelity and robust fitting images across diverse scenarios. These methods often struggle with issues such as texture-aware maintenance and size-aware fitting, which hinder their overall effectiveness. To address these limitations, we propose a novel garment perception enhancement technique, termed FitDiT, designed for high-fidelity virtual try-on using Diffusion Transformers (DiT) allocating more parameters and attention to high-resolution features. First, to further improve texture-aware maintenance, we introduce a garment texture extractor that incorporates garment priors evolution to fine-tune garment feature, facilitating to better capture rich details such as stripes, patterns, and text. Additionally, we introduce frequency-domain learning by customizing a frequency distance loss to enhance high-frequency garment details. To tackle the size-aware fitting issue, we employ a dilated-relaxed mask strategy that adapts to the correct length of garments, preventing the generation of garments that fill the entire mask area during cross-category try-on. Equipped with the above design, FitDiT surpasses all baselines in both qualitative and quantitative evaluations. It excels in producing well-fitting garments with photorealistic and intricate details, while also achieving competitive inference times of 4.57 seconds for a single 1024x768 image after DiT structure slimming, outperforming existing methods.
format Preprint
id arxiv_https___arxiv_org_abs_2411_10499
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle FitDiT: Advancing the Authentic Garment Details for High-fidelity Virtual Try-on
Jiang, Boyuan
Hu, Xiaobin
Luo, Donghao
He, Qingdong
Xu, Chengming
Peng, Jinlong
Zhang, Jiangning
Wang, Chengjie
Wu, Yunsheng
Fu, Yanwei
Computer Vision and Pattern Recognition
Although image-based virtual try-on has made considerable progress, emerging approaches still encounter challenges in producing high-fidelity and robust fitting images across diverse scenarios. These methods often struggle with issues such as texture-aware maintenance and size-aware fitting, which hinder their overall effectiveness. To address these limitations, we propose a novel garment perception enhancement technique, termed FitDiT, designed for high-fidelity virtual try-on using Diffusion Transformers (DiT) allocating more parameters and attention to high-resolution features. First, to further improve texture-aware maintenance, we introduce a garment texture extractor that incorporates garment priors evolution to fine-tune garment feature, facilitating to better capture rich details such as stripes, patterns, and text. Additionally, we introduce frequency-domain learning by customizing a frequency distance loss to enhance high-frequency garment details. To tackle the size-aware fitting issue, we employ a dilated-relaxed mask strategy that adapts to the correct length of garments, preventing the generation of garments that fill the entire mask area during cross-category try-on. Equipped with the above design, FitDiT surpasses all baselines in both qualitative and quantitative evaluations. It excels in producing well-fitting garments with photorealistic and intricate details, while also achieving competitive inference times of 4.57 seconds for a single 1024x768 image after DiT structure slimming, outperforming existing methods.
title FitDiT: Advancing the Authentic Garment Details for High-fidelity Virtual Try-on
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2411.10499