AI-assisted automatic development of RAW photographs.

Three pillars

Three things this does that a RAW developer does not.

Everything below the fold is one of these three, or infrastructure for one of these three. Each pillar states its own mathematics and its own measured result; the eight implementation areas that follow are where those results were taken.

These describe v1.1.0 — the download above. The local description model, the correspondence sidecar, the spatial-tile and free-mask fit stages and the recalibrated retrieval weights all ship in it.

I

AI analysis develop

Your own catalogue is the style reference. Every Lightroom RAW+XMP pair becomes an exemplar carrying a 14-dimension EXIF+histogram feature, the twelve sliders you moved, your tone-curve shape, a colour-family summary and a local-work habit — how many masks you enabled, to which use, at what average strength, never a coordinate. Two local models that bill nothing add a 768-D SigLIP 2 image vector and one Qwen3-VL-2B sentence about the grade. A develop retrieves its 4 nearest past shots by d14 + W_EMB(1−cos) + W_TXT·z(1−cos−hub) + W_DESC·z(1−cos), where hub subtracts each exemplar's mean vocabulary cosine so a photo that scores high against every direction stops hoarding the top-4; W_EMB=4, W_TXT=0.5, W_DESC=0.5 are the leave-one-out winners over the real corpus after that correction, MAE 0.713143 → 0.688864 with a paired 95% CI of [+0.005837, +0.041111] — and opposite direction texts now retrieve different photographs (antonym top-1 overlap 71% → 44.7%, 149 of 169 exemplars reachable). Their habits reach the model behind an untrusted-data fence, and a capped pull — 0.18 at the shipped Style 0.3, full at 1.0 — moves the proposal toward your means without copying one.

Details: AI advisor and reverse fit →

II

AI full-image generation, recovered as an editable recipe

reimagine composes every prompt onto an unconditional faithfulness scaffold, because input_fidelity is negotiated away by gpt-image-2, then measures its own result with the same statistic the fit uses: D = sqrt((1−rho_grad)² + E_pyr²) over a shared 384×256 raster. Below D = 0.35 the full solve runs; at or above it a bounded Atmosphere mode caps EV at ±1, WB gains to [0.80, 1.25], saturation to ±30 and curve slope to [0.5, 1.5]. The global tone map is a 64-bin Tukey-biweight IRLS regression (c = 4.685, three rounds from a weighted median), so invented content loses weight by the estimator's own influence function; a per-band colour mixer then solves saturation and luminance one hue band at a time from that band's own population — hue rotation never — and is judged twice against its own absence. Local work attaches in one fixed order — semantic regions or luminance bands, a frozen-evidence quadtree to 4×4, then at most two free-form remainder masks — each through the same 3% two-sided evidence, D < 0.65, CI-excludes-zero and 0.012 rim gates, and a frame-regression budget that is deliberately not uniform — 0.02 for a semantic region, zero for a luminance band, a tile or a free mask. What you get back is a recipe you can edit and export, never the generated pixels.

Details: layered spatial reverse-fit →

III

The mathematics of matching Lightroom

Camera Raw's own scope is read and written by a hand-rolled XML layer, so unknown namespaces survive a merge and sixteen real radial sidecars round-trip their four corners byte for byte; measured domains include LocalExposure2012 = EV/4, local Hue /180 and global Sharpness 1:1, and a raster mask is reported as a named bitmap loss rather than approximated. The pixel side is fitted the same way: a monotone Fritsch–Carlson tone LUT, a two-arm Texture model (A1=0.172443, A2=0.304888, all 45 of 45 Lightroom anchors within ±0.02), a measured 290×11 radial feather table, and the brush law k(rho;h) = (1 − rhom(h))n(h) with a one-parameter flow law kappa = 0.1284 (held-out RMS 0.0109). Mask coordinates follow measured frame laws: radials cross m_lr−1 ∘ T_engine exactly once on a zero-parameter centre raw_full_dims/2 − DefaultCropOrigin and close 41/41 vectors to ≤1 px, while linear handles are transported to keep the line straight and are openly reported as not pixel-closed at 9.748/7.025/6.336 px RMS. The falloff is the C1 smoothstep t²(3−2t): a free-end fit reads RMS 0.0045 against 0.017 for a straight ramp.

Details: Lightroom mask-frame laws →

Tech stack & algorithms

The implementation details—not a logo strip.

Eight implementation areas connect decode, measured rendering, local and AI selection, coordinate transport, sidecars, AI proposal, and application infrastructure. Each Details link opens the canonical equations, parameter provenance, measurements, disclosures, and source paths.