RFC-GAN — Feedback-Conditioned GAN for Iterative Memory-to-Sketch Facial Translation

Kareem Hassani, Kevin Abi Zeid Daou, Roni Bou Saab, Hadi Al Mubasher, Mariette Awad

Keywords: Generative Adversarial Networks (GANs), Feedback-conditioned learning, Iterative refinement, Self-Attention, Edge-Gating, Image-to-image translation

Abstract: Eyewitness–to–sketch translation is traditionally performed by human artists, a process prone to bias and information loss. Recent work has applied conditional GANs to automate this task, yet existing models remain limited in their ability to iteratively refine coarse recollections into photorealistic faces. We propose a context-feedback training paradigm for image-to-image GAN: at each step, the generator and discriminator receive the most recent output as an auxiliary three-channel input, enabling the model to reason over its own predictions. Building on the Pix2Pix framework, we further investigate where to embed Self-Attention and Edge-Gating modules within the encoder–decoder and skip connections, systematically analyzing their effect on perceptual and adversarial loss. Experiments on synthetic-sketch and real eyewitness datasets demonstrate consistent improvements in FID, LPIPS, and human-rated realism, with the model producing sharper, more faithful faces and exhibiting stable iterative refinement at inference. These results suggest that feedback-conditioned GAN provide a principled path toward reliable facial reconstruction from memory.

Primary Area: generative models

Submission Number: 24356