USDM - Unsupervised Single-Image Face Demorphing via Reconstruction-Guided Latent Optimization
Nitish Shukla, Rajeev Dwivedi, Vinod Kumar Kurmi, Arun Ross; Under review
Research Goal
Existing demorphing methods train on sensitive biometric data and often need retraining for each new domain. USDM recovers both contributors of a morph with no training at all: no morph–contributor pairs, and no finetuning of any network.
How it works
- Start from the morph. A StyleGAN inversion of the input morph gives a starting latent code.
- Optimize two residuals. Two latent residuals are optimized jointly so that the two generated faces, re-morphed, reproduce the input.
- Identity-aware objectives. Alongside reconstruction, identity-consistency and contributor-separation terms keep the two outputs distinct and identity-preserving. The encoder, generator and face-recognition network stay frozen.
Status: under review.