What it does

Exposure is a white-box photo post-processing framework that trains a reinforcement-learning agent to apply sequences of differentiable filters (exposure, contrast, tone curve, etc.) to RAW images. It learns from unpaired datasets—no need for aligned before/after pairs—using a GAN-based critic and a reward signal that favors natural-looking outputs. The system outputs not just a retouched image, but the actual sequence of editing steps that produced it.

The interesting bit

The “white-box” angle is the hook: unlike black-box neural filters, this agent produces human-interpretable editing actions. The authors even built a UI to have humans play the same RL game, collecting ground-truth trajectories for comparison. The “infinite resolution” claim comes from operating on filter parameters rather than full-resolution feature maps—though the README is vague on whether this holds in practice for arbitrarily large images.

Key highlights

• Trains in ~100 minutes on a GTX 1080 Ti using the MIT-Adobe FiveK dataset (~2.4GB download)
• Pretrained model included as a git submodule; clone with --recursive
• Designed for linear 12+ bit RAW workflow; 8-bit JPEG/PNG inputs risk posterization and color-space mismatch
• One-to-many mapping via dropout means results vary across runs—intentionally, not a bug
• SIGGRAPH 2018 paper with slides, video, and recorded human-editing tutorial

Caveats

• Explicitly labeled a research prototype: “a lot of engineering efforts required to make it suitable for a real product”
• 8-bit inputs work “to some extent” but can produce posterization artifacts; sRGB gamma mismatch degrades quality
• Some pretrained models (artist-specific ones) withheld for copyright reasons

Verdict

Worth a look for computational photography researchers and anyone building interpretable, editable image pipelines. Skip it if you need a production-ready Lightroom plugin—the authors are upfront that this isn’t one.