Imagen — learning a photographer’s edit from RAW brackets
Taught a neural network to edit real-estate photos the way one photographer does, for a French photography agency.
My partPipeline author, with Layth Ayache on training
In short
A freelance pipeline for a French real-estate photography agency that ingests bracketed RAW sessions, decodes them through a color-accurate DNG pipeline into 16-bit linear frames, pairs them with the photographer’s edited JPEGs, and trains an HDRNet model to replicate the editing style.
- What it is
- A dataset pipeline and training setup that teaches a neural network to edit real-estate photographs the way one photographer does.
- Why it matters
- Architectural photographers shoot ambient brackets, flash fills, and lights-on stacks for every frame, then hand-blend and grade them. The style is consistent but slow to reproduce.
- My part
- I designed and built the pipeline and its color science; Layth Ayache contributed training and delivery runs.
Step by step
- 1
Ingest
Detect −6/−3/0/+3/+6 EV ambient brackets, flash fills, and lights-on stacks across any listing folder structure; write an index per listing.
- 2
Decode
Sony ARW through LibRaw; Canon EOS R5 DNG 1.7 with JPEG XL tiles through a tifffile fallback and the DNG specification’s color pipeline.
- 3
Target
The photographer’s JPEG is linearized so the artistic values are preserved exactly.
- 4
Train
HDRNet learns the mapping; the pipeline resumes automatically after interruption.
Results
- 2–3 photos/day
- Editing throughput before
- a team of three, by the client’s count
- ~40 photos/day
- Editing throughput after
- per team member, by the client’s count
For engineersHow it is built, its limits, and its keywords.
How it works
Brackets are grouped by timestamp; RAW files are decoded with the full DNG color pipeline (linearization, white balance and camera calibration, forward matrix, Bradford adaptation) and lens correction into 16-bit linear sRGB; the edited JPEG is linearized as the target; an HDRNet bilateral-grid network is trained on the pairs.
Limits
- The photographs belong to the client, so none appear here, and the code is private.
- Style transfer is per photographer; the model does not generalize across agencies.
Keywords
- HDRNet
- RAW processing
- DNG
- color pipeline
- Python
- PyTorch