This project investigates the intersection of multi-dataset training, domain shifts, and generalization in the
context of image-to-image diffusion models for Virtual Try-Off (VTOFF) task.
It aims to understand how training on diverse datasets affects the performance and robustness of such models
across different domains.
Key research questions include:
- What is the effect of training image-to-image diffusion models for the VTOFF task on domain-specific datasets?
- Does training on multiple datasets improve the generalization capabilities of VTOFF models across domains?
- How do domain shifts between datasets (e.g., clothing styles, poses, image quality) impact model performance?
Example tasks include:
- Train a VTOFF model on VITON-HD and evaluate it on Dress Code, and vice versa,
- Train a model on both datasets and compare results with single-dataset results,
- Analyze the qualitative and quantitative differences in model outputs to assess the impact of domain-specific and multi-domain training.
Literature