Reproduce “diffDemorph: Extending Reference-Free Demorphing to Unseen Faces” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nitish Shukla, Arun Ross
diffDeMorph, a novel reference-free facial demorphing method from Michigan State University, effectively disentangles two constituent face images from a single morph image. It successfully generalizes across diverse morphing techniques and face styles by training on synthetic data and testing on real-world morphs, achieving substantial improvements in biometric matching rates, including up to a 68.78% higher True Match Rate on the StyleGAN dataset.
Key insights
- Direct RGB domain conditioning of the diffusion model is critical for effective demorphing, as it preserves perceptually important details (e.g., hair, background) that latent space compression might discard.
- Formulating demorphing as a coupled image generation task, where two constituent images are recovered as a single entity, significantly improves disentanglement and visual fidelity.
- Training on synthetic morphs and evaluating on diverse real-world morphs demonstrates strong generalization capabilities, addressing data privacy and scarcity concerns while maintaining performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.