Teach a pre-trained AI model to edit photos like you, right from your browser. This locally-hosted data app allows you to easily ingest Lightroom catalog files, train and save models, and generate predictive Lightroom develop settings and/org XMP data for images.
- Python 3.9+
- Exempi 2.2.0+ (for Python XMP Toolkit) - installed by start script
Run the start scrpt to initialize venv, install all dependencies, start the app, and open the UI:
# UNIX-like environments
./run.sh
# Cross-platform Python launcher
python3 run.py
Drag-and-drop as many Lightroom catalog .lrcat files into the UI as you want. Click "Run Ingest" to extract image metadata, decode Adobe XMP data, and parse internal develop settings.
Can be run from the command line as well:
# Activate venv if not already
source .venv/bin/activate
python -m modules.ingest \
--catalog /data/raw/catalog.lrcat \
--previews /data/previews \
--out_csv /data/models \
--criteria "flag=flagged,colorlabel=green"
Train and save models using previously ingested slider datasets in .csv format. References unedited image previews to learn ideal slider settings based on raw image data.
Can be run from the command line as well:
# Activate venv if not already
source .venv/bin/activate
python -m modules.train \
--csv /data/dataset/sliders.csv \
--previews /data/previews \
--out_model /data/models/model.pt \
--epochs 5 \
--batch_size 32
Use model inference to predict slider values given image file inputs. The goal is for the trained model to independently generate XMP side-car files and/or internal Lightroom develop settings from image inputs, so that when you open Lightroom, all your images are already edited in your style.
Using Lightroom-SQL-tools to handle extracting data from .lrcat files.
See the full .lrcat table schema here, and an example XMP file here
From my digging, the tables of interest are:
Adobe_AdditionalMetadata
- Column: xmp (TEXT) — contains the full XMP side-car XML for each image.
Adobe_imageDevelopSettings
- Columns: numeric fields like grayscale, hasPointColor. You can still pick up a handful of basic flags here, but the heavy lifting lives in the text column, which is a Lua-like data objecet with essentially all the develop settings that would be in the Camera Raw Settings namespace of the XMP. Unfortunately, the CRS attributes are not always included in the XMP dta returned from the catalog (this is a setting in Lightroom). That makes this table more reliable than going to the XMP. Unfortunately^2, Lightroom-SQL-tools doesn't currently support this data yet, so we have to get it ourselves with sqlite3.
AgHarvestedExifMetadata
- Columns: aperture, shutterSpeed, isoSpeedRating, cameraModelRef, dateYear/dateMonth/dateDay, etc.
Run pytest to run all tests under tests/