Computer Animation and Virtual Worlds Программное обеспечение
Efficient Parser‐Free Virtual Try‐On With Context‐Aware Alignment and Edge‐Guided Detail Generation
Аннотация
ABSTRACT Virtual try‐on (VTON) is an important technique for human‐centric garment visualization, digital fashion presentation, and virtual human appearance editing. Existing methods often rely on explicit human representations, such as pose estimation and semantic parsing, which require additional preprocessing, prolong the inference process, and make the synthesis quality sensitive to structural estimation errors. To address these limitations, we propose CAEG‐VTON, an efficient parser‐free VTON framework with context‐aware alignment and edge‐guided detail generation. CAEG‐VTON adopts a teacher–student paradigm, where ContextEdgeNet uses rich human representations during training to learn garment–body alignment and detail‐preserving synthesis, while LiteFitNet inherits such knowledge through distillation and performs inference using only the human image and target garment. LiteFitNet further integrates compact feature extraction, human–garment feature fusion, and an inverted‐residual generation structure to efficiently preserve garment shape, boundaries, and texture details. Experiments on VITON and VITON‐HD demonstrate that CAEG‐VTON improves garment structure, boundary clarity, texture preservation, and distribution‐level realism while maintaining a favorable balance between model compactness and inference efficiency. These results show its potential for lightweight virtual garment visualization.