Members: Jiya Sinha, Rashi Bharti, Sakshi Tiwari, Nandhana KS
This repo contains code for out Region-Aware Makeup Transfer using GCNs.
Mean score: 4.9/10
Mean score of images by 54 users:
XFlow-GAT was the first architecture we experimented with. The model followed a multi-stage pipeline: node features → graph message passing → updated node features → rasterization → U-Net → image. The idea was to let a GAT capture fine-grained interactions between corresponding facial landmarks and then decode those refined node embeddings into a full image. In practice, this design turned out to be unnecessarily complicated for the relatively constrained task of region-based makeup transfer. The model hallucinated unnatural colors, oversmoothed key regions, and failed to maintain global consistency. The figure below shows an example of the output produced by XGAT-Color. The overly smooth eyelids, distorted lip coloration, and color bleeding into the background demonstrate the instability of this approach when used for direct makeup transfer.
