Constraint-based visual control
Give visual intent
a coordinate it can reach.
Move the center of chromatic emphasis. Keep the original hue-family totals, lightness and a protected region. The editor solves the change—or shows why these controls cannot make it.
22 / 42requests executed and verified in 16-bit
20 / 42requests declined with a verified bound
< 0.0016%worst hue-family budget drift, executed edits
6 paintings3 primary + 3 further validation images
These are bounded color-control results. No objects move, no new content is generated, and this is not a generative-model leaderboard. Six 8-bit export failures are retained below.
Original
Source pixels, normalized once to a 512-pixel maximum dimension.
Coordinate-only adjustment
Constrained adjustment
Loading recorded experiments…
White dot: original chroma centroid. Green dot: measured result. Amber cross: requested target. Dashed box: unchanged central region. Markers are annotations, not edits to the saved images. Browser display may reduce the 16-bit masters to display precision.
The surprise: undo can be infeasible.
In the Degas image, right + down is reachable under the preservation locks. Down alone is not. Removing the horizontal decision while preserving the vertical one conflicts with the remaining constraints. This editor detects that conflict before committing the change.
This is a dependency in the bounded color-adjustment family, not a claim that all possible editing techniques face the same impossibility.
What the comparison actually establishes
Both methods use identical spatial color controls. The coordinate-only version optimizes the target without hue-budget constraints. The constrained version adds six chroma-total locks. This is an ablation of constraint enforcement, not a comparison against an expert colorist.
| Across six images | Coordinate-only | Constrained |
|---|
| Requested edits | 42 | 42 |
| Computed edits | 42 | 22 |
| Verified full contract · 16-bit | 0 | 22 |
| Declined as infeasible | 0 | 20 |
| Verified full contract · original 8-bit run | 0 | 16 |
The coordinate-only version hits all 42 targets in 16-bit, but violates the hue budgets. The constrained version declines 20 targets and verifies all 22 executed masters. Six additional all-controls-locked contradictions are correctly declined by each method.
Export failures and why precision matters
The primary 8-bit run passed all eight computed candidate edits. On three additional images, only eight of fourteen computed candidates passed: five nearly monochrome Jeong Seon edits missed their chroma coordinates after rounding, and one Klimt edit exceeded the lightness tolerance. Those failures remain in the original records.
A post-hoc check exported the exact same parameters at 16-bit without re-optimization. All 22 computed candidates passed. Tiny chroma signals in a nearly monochrome reproduction are not necessarily meaningful artistic decisions; precise control of such a signal is not proof of perceptual value.
Precision records · Original additional-image results
Scope, implementation and prior work
The engine reuses the user's Color Kernel conversion functions and ISR field readouts, with SciPy HiGHS and SLSQP for feasibility and optimization. It changes chromatic emphasis at fixed pixel locations and fixed OKLab lightness. It does not move objects or separate material from lighting. Hue budgets mean six source hue-bin chroma sums; they do not preserve the entire color distribution.
Coordinate-based image optimization is established: DiffVG and DragDiffusion are relevant precedents. OKLab supplies the color geometry. This work tests an executable connection between the supplied measurements, bounded controls, feasibility witnesses and verified export. Novelty and artist-time savings remain unproven.