Skin Tone Transfer in the Wild: A Comprehensive Comparison of Colour Transfer Approaches

Skin tone transfer involves adjusting the skin tone in an image to match that of a reference while preserving perceptual realism and individual identity. Manual editing for this task is labour‐intensive and requires substantial expertise, motivating the need for automated approaches. As a subtask of example‐based image transfer, skin tone transfer can benefit from existing methods developed for colour and style adaptation. We present a benchmark of nine representative methods spanning classical and learning‐based colour transfer, photorealistic style transfer and GAN‐/diffusion‐based image translation. Using a curated in‐the‐wild portrait set, we evaluate outputs with objective metrics, skin‐region colour‐difference measures and a controlled psychophysical study with human ratings. We find that widely used metrics correlate weakly with perceptual preference, making metric‐only evaluation unreliable. Qualitative analysis further shows that large tone gaps or substantial differences in the skin‐region intensity distributions between source and reference more often lead to unnatural results across methods.