Thermal Face Recognition From Synthetic Data: FLUX.1 Kontext Image-to-Image Adaptation and Cross-Sensor Evaluation

Abstract

Thermal face recognition (FR) is well suited to surveillance under nighttime and adverse illumination, but its progress is constrained by the scarcity of large, annotated thermal face datasets. This work investigates whether a generative foundation model can be adapted with very few samples to produce identity-preserving thermal faces from visible images, and whether such synthetic data can train recognizers that remain effective on real thermal imagery. We fine-tune FLUX.1 Kontext, a flow-matching model with explicit image-to-image (I2I) conditioning, using Low-Rank Adaptation (LoRA) on as few as 50 synchronized visible–thermal pairs: the visible image anchors facial geometry and identity, while a fixed text prompt guides thermal appearance. With the selected 50-pair configuration we generate PUCV-TER-S, a synthetic thermal dataset of 7,800 images spanning 78 identities. Under an identical data budget, the adapted model preserves subject identity where adversarial baselines fail: 99.9% rank-1 identity preservation, versus 11.4% for Pix2Pix and 3.5% for CycleGAN. An ArcFace-based recognizer trained exclusively on the synthetic data is evaluated on real thermal benchmarks, distinguishing a controlled same-identity setting (PUCV-VTF-SET), which saturates and is therefore read as evidence of synthetic-to-real transfer, from two more demanding unseen-identity cross-sensor settings (SpeakingFaces and UXX Thermal Faces), where it attains high Top-1 accuracy with competitive verification performance (ROC-AUC, EER, and TAR@FAR), all reported with 95% confidence intervals over repeated gallery/query resampling. The full pipeline runs under fixed random seeds, and a from-scratch re-execution quantifies its run-to-run variance while confirming every conclusion, supporting synthetic thermal data as a practical training source for thermal FR under real, cross-sensor conditions.