Semi-Supervised Parametric Real-World Image Harmonization

Composite IHT Harmonizer Ours Parametric curves & shading map Fully supervised (previous work) Dual-stream semi-supervised (Ours) Figure 1. Visual comparisons between state-of-the-art harmonization methods IHT [9], Harmonizer [14], and ours. Our model is fully parametric. This gives artists full posterior control over the final composite, makes runtime efficient for high-resolution real-world inputs and regularizes training. Our model predicts global RGB curves and a local shading map (right). Benefiting from the novel dualstream semi-supervised training strategy, our method (right) produces more realistic harmonized images on real-world composites (left). This new training strategy, together with the shading map, makes it the first harmonization method to address local tonal adjustments, such as shading the face according to the sun's direction (top) or selectively darkening the part of the dog inside the cave (bottom).