Workshop #1: First Runs
Late August to early September 2026. Setup, renderer, three training runs, and what they showed. Technical notes from the machinery side of the project; the narrative is in Journal #4. Data Corpus: 1,414 raw files with the photographer’s Lightroom develop settings (XMP), harvested through the Lightroom partner API. Cameras: Fujifilm X100VI, X-E5, X-T5; Ricoh GR III; Leica D-Lux 8. 203 of the 1,414 are matched to published Instagram posts and carry a published tier label; the rest are edited. (The catalog as a whole has 219 published matches; 16 are JPEG originals outside this corpus.) Labels: the develop XMP is the label. It carries global tone (exposure, contrast, highlights, shadows, whites, blacks, texture, clarity, dehaze), point curves, HSL, colour grading, calibration primaries, grain and vignette, crop and rotation, upright/perspective, the camera profile or film-simulation Look, and parametric masks (linear and radial gradients with their local adjustments). Renderable subset: 870 of the 1,414. Excluded: 540 edits whose masks were drawn with Lightroom’s AI selection tools (the mask rasters are computed on the desktop client and are not available to the cloud engine), and the 9 Leica D-Lux 8 DNGs, which the cloud engine does not process. Neutral inputs: each raw is also rendered once with no develop settings (camera DCP colour profile, lens distortion correction only) at 512 px, as the model’s input image and as the reference for palette measurement. Splits: date-blocked (whole capture dates go to one side, so near-duplicate frames stay together). Imitation: 783 train / 53 valid over the 856 supported labels, holdout frames excluded (71 of the train rows are published tier). Curator: 9,870 pairs, 494 held out. Crop probes: 1,243 / 135 over all 1,414 labels; pairwise crop set: 2,638 / 327 pairs on the same split. Renderer Two renderers were used. ...