dc-GAN - Dual Conditioned GAN for Face Demorphing From a Single Morph

Nitish Shukla, Arun Ross; In Proceedings of 19th IEEE FG 2025

Research Goal

Existing demorphers assume some identities are shared between train and test sets, and tend to merely replicate the morph. This work removes the shared-identity assumption and is the first to overcome morph-replication when test morphs use entirely unseen identities (the hardest “scenario 3”), in both reference-free and reference-based settings.

How it works

  • Dual conditioning. A UNet generator is conditioned on (i) the morph image in the image domain and (ii) the morph embedding $\mathcal{E}(x)$ injected into its latent layers — giving far richer guidance than the morph image alone.
  • Conditioned discriminator. The discriminator also sees the morph, distinguishing a real triplet (morph, BF1, BF2) from a fake one (morph, OUT1, OUT2). The two conditions work in tandem to suppress morph-replication.
  • Cross-road loss aligns the unordered generator outputs with the ordered ground-truth faces. For reference-based (differential) demorphing, the encoder simply takes the morph and a reference image (6-channel input).
An image encoder embeds the morph to condition a UNet generator (alongside the morph image); the discriminator separates real from synthetic (morph, face, face) triplets.

Key results

  • 93.86% average TMR on AMSL (ArcFace), +23.32% over the prior GAN baseline; strong TMR across OpenCV (93.99%), WebMorph (89.87%), FaceMorpher (94.39%), MorDiff (93.75%).
  • Wins decisively in scenario 3 (unseen identities) while staying competitive in scenario 1.
  • Ablations: removing the cross-road loss drops performance −20.87%; removing the embedding condition $\mathcal{E}$ drops it −8.57% — both conditions matter.
  • Reference-based (differential) demorphing reaches 94.88% TMR on AMSL.

Resources

📑 Paper 💻 Code

Citation

If you use this work, please cite:

@INPROCEEDINGS{shukla2025dcgan,
  author={Shukla, Nitish and Ross, Arun},
  booktitle={2025 IEEE 19th International Conference on Automatic Face and Gesture Recognition (FG)},
  title={dc-GAN: Dual-Conditioned GAN for Face Demorphing From a Single Morph},
  year={2025},
  pages={1-9},
  doi={10.1109/FG61629.2025.11099072}}