Let an AI agent rescue your backlit raw photos

I photographed a family celebrating their new roof, shooting up at them against a bright afternoon sky. The camera JPEGs were silhouettes. Instead of opening a raw editor, I gave Claude Code this prompt:

there is a number of raw images taken yesterday upon /Volumes/Untitled/DCIM/100MSDCF/*.ARW, since they were taken in the harsh light, can you use the Sony raw images to better the image and save the final outputs as AVIF for sharing? I took several so please whittle down the family pictures celebrating their new roof to a handful please.

Minutes later I had five good photos.

What it did

1. Found “yesterday” by reading the capture dates with exiftool, which left 27 of the card’s 50 frames.
2. Made a contact sheet from the JPEG previews embedded in the ARW files. This split the frames into family bursts and shots of just the building.
3. Culled each burst by decoding the raws with the shadows lifted and comparing faces at 100% (open eyes, smiles, nobody hidden behind a scaffold pole). It kept the best frame from each of 5 bursts.
4. Developed the raws using exposure fusion (details below).
5. Saved AVIFs at full resolution, around 1.1–1.9 MB each, then copied the capture date and camera/lens tags back from the ARWs. It also checked for GPS before sharing; there was none.

The development trick

A raw file holds several stops more than the JPEG shows. The agent decoded each ARW to linear 16-bit with rawpy (LibRaw), then made synthetic exposures from −1 to +3 EV out of that one file. It blended them with OpenCV’s Mertens exposure fusion, which keeps the well-exposed parts of each version:

base = 0.95 / np.percentile(luminance, 99.7)  # brightest sky just under clipping
stack = [to_srgb(linear * base * 2.0 ** ev) for ev in (-1, 0, 1, 2, 3)]
fused = cv2.createMergeMertens().process(stack)

After that came a light finish: set the black and white points, add a gentle S-curve, a shadow lift on the backlit frames and some vibrance to bring the sky back to blue. Pillow 12 writes AVIF natively, so no extra encoder was needed:

uv run --with rawpy --with opencv-python-headless --with pillow python develop.py

Why this works well as an AI task

• It looks at the photos. The agent reviewed contact sheets and face crops itself, so the cull was based on expressions, not filenames.
• It checks its own work. It compared before and after, noticed its sky was paler than the camera’s, and fixed it in a second pass.
• No editing software. LibRaw, OpenCV and Pillow are installed on demand with uv, so there’s no Lightroom catalogue and nothing to tidy up afterwards.

The tip: when you have a card full of difficult raws, describe the result you want, not the slider settings. Ask the agent to show its picks, and you get a curated, shareable set.