yuanming-hu/exposure

Combines differentiable photo editing operations (exposure, saturation, shadows, highlights) with reinforcement learning to discover optimal editing sequences, trained on paired expert retouching data from the MIT-Adobe FiveK dataset. Built on TensorFlow with support for high-bit-depth RAW images and linear color spaces, enabling interpretable editing chains rather than black-box neural transformations. Handles one-to-many mappings through dropout-based stochasticity, allowing diverse stylistic outputs from single inputs.

780 stars. No commits in the last 6 months.