diffDeMorph - Extending Reference-Free Demorphing to Unseen Faces

Nitish Shukla, Arun Ross; In Proceedings of IEEE ICIP 2025

Research Goal

Earlier reference-free demorphers assume the test morphs use the same morphing technique and same face style as training. This work removes those assumptions: a demorpher trained purely on synthetic morphs that generalizes to real morphs made with unseen techniques and styles — the realistic operational setting.

How it works

  • Coupled forward diffusion. The two faces are treated as a single 6-channel object $i=(i_1,i_2)$ and noised together, so the reverse process must denoise both at once rather than producing two near-duplicate marginals.
  • Morph-guided reverse sampling in RGB. At each step the denoising UNet (9-in / 6-out) concatenates the morph image in the RGB domain with the noisy sample. Conditioning in pixel space — rather than a compressed latent — lets the model attend to hair, background, and other low-salience details essential for demorphing.
  • This fixes a key flaw in prior diffusion demorphing, whose evaluation lacked a true backward-sampling step (reducing it to an autoencoder) and so failed on unseen morphs.
Coupled diffusion: both constituent faces are reconstructed jointly from noise, with the morph injected as RGB guidance at every denoising step.

Key results

  • First method to generalize across morph techniques and face styles; beats the prior state of the art by ≥59.46% Restoration Accuracy under a common protocol (per-dataset RA gains of 73.80 / 59.46 / 62.18 / 74.21 / 61.88 / 79.75%).
  • TMR@10%FMR gains of 27–69% across the six datasets over the closest competitor; near-saturated TMR (99–100%) on AMSL, OpenCV, FaceMorpher, WebMorph, MorDiff, StyleGAN.
  • RGB vs. latent ablation: RGB conditioning reaches 99.21% TMR / 99.97% RA vs. 87.72% / 97.44% in latent space — latent compression erases details demorphing depends on.

Resources

📑 Paper

Results

(Left) Demorphing across multiple techniques and styles. (Right) RGB conditioning preserves hair/background that latent compression removes.

Citation

If you use this work, please cite:

@inproceedings{shukla2025diffdemorph,
  title={diffDeMorph: Extending Reference-Free Demorphing to Unseen Faces},
  author={Shukla, Nitish and Ross, Arun},
  booktitle={IEEE International Conference on Image Processing (ICIP)},
  year={2025}
}