Crosstalk From Datasheet

Crosstalk from datasheet

Overview

A color negative's three dye layers each leak density into channels they shouldn't ("unwanted absorption"), muddying color. NegPy's Crosstalk control (Process panel) unmixes them with a 3×3 matrix applied to the raw negative densities. This skill turns a film's published spectral dye-density curves into that matrix: read the unwanted absorptions, invert them, write a .toml.

Core principle: the matrix is the inverse of the dye-leakage matrix. You only read values off curves and judge data quality — derive_matrix.py does the arithmetic.

Read docs/CROSSTALK.md and crosstalk/README.md once for the file format and the density-domain math NegPy applies (d=-log10(rgb_negative), d_out=M·d, pre-normalization); NegPy row-normalizes M and blends with identity by the Crosstalk strength (0–1), so only the off-diagonal/diagonal ratios matter — absolute row scale is irrelevant.

Workflow

1. Acquire the source.

   • Local PDF → Read it with pages (vision — the curves are pixels; a text fetch will not see them).
   • URL → PDF → curl it to the scratchpad, then Read the pages as images.
   • HTML page → WebFetch for any numeric spectral tables; if it only embeds a plot, download the image and Read it.
   • Raw spectral numbers / CSV → use them directly (Tier 1).

2. Classify the data tier (the one decision that's easy to get wrong):

   digraph tier {
     "Separated C/M/Y dye-density curves\n(or numeric spectral data)?" [shape=diamond];
     "Only aggregate neutral dye-density\n(one combined curve)?" [shape=diamond];
     "Tier 1: derive directly" [shape=box];
     "Tier 2: estimate, label (approx)" [shape=box];
     "Tier 3: do NOT fabricate" [shape=box];
     "Separated C/M/Y dye-density curves\n(or numeric spectral data)?" -> "Tier 1: derive directly" [label="yes"];
     "Separated C/M/Y dye-density curves\n(or numeric spectral data)?" -> "Only aggregate neutral dye-density\n(one combined curve)?" [label="no"];
     "Only aggregate neutral dye-density\n(one combined curve)?" -> "Tier 2: estimate, label (approx)" [label="yes"];
     "Only aggregate neutral dye-density\n(one combined curve)?" -> "Tier 3: do NOT fabricate" [label="no"];
   }
   

   Most consumer datasheets (e.g. Kodak Gold 200) are Tier 2: the spectral-dye-density plot is one aggregate "midscale neutral" curve, not separated dyes. Spectral-sensitivity curves are not dye-density and must never be used as a substitute.

3. Read the leakage values into matrix A (rows = R/G/B measurement bands, cols = C/M/Y dyes). Dye→channel: cyan↔Red, magenta↔Green, yellow↔Blue. Sample each dye's diffuse density at band centers R≈650 / G≈550 / B≈450 nm (datasheets use Status M). Each dye's own-band value is the diagonal; the other two are its unwanted absorption (positive). Reading a graph by eye is ±~0.05 density — fine, the result gets row-normalized.

   • Tier 1: read all nine values off the separated curves.
   • Tier 2: diagonal = own band; estimate the off-diagonals from the visible secondary humps in the aggregate curve plus known dye chemistry (magenta's unwanted blue and some red; cyan a little green/blue; yellow nearly clean). Name it "… (approx)".
   • Tier 3: stop. State that the datasheet lacks dye-density data. Optionally emit a conservative starter close to NegPy's Default, named "… (approx, starter)", and suggest chart-based calibration instead. Never invent precision.

4. Run the deriver:

   echo '{"name":"Kodak Gold 200 (approx)","readings":[[1.0,0.06,0.02],[0.04,1.0,0.05],[0.02,0.12,1.0]]}' \
     | python3 .claude/skills/crosstalk-from-datasheet/derive_matrix.py --out <film>.toml
   

   readings is A row-major (R/G/B band rows). It column-normalizes, inverts, and rescales so the diagonal reads 1.0; it exits non-zero on a singular or misshaped input rather than emitting a bad matrix.

5. Review & self-validate — never trust the loader's silence (it skips malformed files without error):

   • 3×3 numeric, no booleans.
   • Diagonal 1.0; off-diagonals negative (a correction subtracts contamination) and small — flag anything |·| > 0.2 as a likely misread/transpose.
   • Confirm it actually loads:

     uv run python -c "from negpy.services.assets.crosstalk import CrosstalkProfiles as C; \
       r=C._parse_file('<film>.toml'); assert r, 'INVALID — would be skipped'; print('ok', r[0])"
     

6. Write the file — pick the destination by intent:

   • Contribute / bundle → repo gallery crosstalk/<film>.toml.
   • Personal use now → <Documents>/NegPy/crosstalk/<film>.toml (resolve the Documents path via get_default_user_dir() in negpy/kernel/system/paths.py).

Quick reference

Common mistakes

• Transposing rows/cols — A is bands×dyes; the output is out-channels×in-channels. Keep R/G/B order throughout.
• Positive off-diagonals in the output — wrong sign; a correction must subtract leakage.
• Using sensitivity curves as dye density — they measure different things; not interchangeable.
• Over-claiming precision — an aggregate (Tier 2) curve gives an estimate; say so in the name.
• Forgetting it's density-domain — read density off the dye-density plot, not transmittance.

Worked example — Kodak Gold 200 (Tier 2)

Read the datasheet PDF pages as images. The CURVES section has Characteristic, Spectral- Sensitivity, and Spectral-Dye-Density plots — but the dye-density plot is a single aggregate "midscale neutral" curve → Tier 2. Diagonals = 1.0; estimate off-diagonals from the curve's secondary absorptions + dye chemistry (magenta's unwanted blue is the biggest term):

A (rows R/G/B, cols C/M/Y):   R:[1.0,  0.06, 0.02]
                              G:[0.04, 1.0,  0.05]
                              B:[0.02, 0.12, 1.0 ]

Run the deriver → emitted TOML (note name flags it approximate):

name = "Kodak Gold 200 (approx)"
matrix = [
  [1.0000,  -0.0579,  -0.0171],
  [-0.0390,   1.0000,  -0.0492],
  [-0.0152,  -0.1191,   1.0000],
]

The strong B,−0.119(G) term is the magenta-dye unwanted-blue correction. Validate it loads, then write it to crosstalk/kodak_gold_200.toml (contribute) or the Documents folder (personal).