Spectral Intuition and Color Shift

1. Key Premise Clarification

Film Simulation ≠ Color Preservation Algorithm

Film simulation typically allows for overall tonal changes (such as warm tones, cool tones, curve compression), but does NOT allow:

Unexpected channel crosstalk (e.g., Red being asymmetrically pulled)
Neutral colors becoming non-neutral
Uncontrollable hue shifts related to luminance

So the goal is not "Output = Input", but rather:

Verify if there are abnormal color shifts within the "known acceptable range of variation"

2. Core Idea: Use "Diagnostic Images"

Natural photos cannot be used for algorithm verification because you can never prove whether the color cast comes from the algorithm or the content itself.

You need Synthetic / Diagnostic Images.

3. Category 1: Neutral Axis Verification (Most Important)

1️⃣ Input: Ideal Neutral Grayscale

Construct an image:

R = G = B
From 0 → 255 (or linear 0 → 1)
Include different luminance sections (Shadows / Mid-gray / Highlights)

Expected Behavior

After film simulation:
- R′ ≈ G′ ≈ B′
- Overall brightening/darkening/curve changes are allowed
- Systematic deviation of Δ(R′−G′) with luminance is NOT allowed

Calculable Metrics

Metric A: Neutral Color Deviation

Δ_neutral = mean(|R' - G'| + |G' - B'| + |B' - R'|)

Metric B: Luminance-Correlated Deviation

corr(L_in, (R'-G'))
corr(L_in, (G'-B'))

If color cast changes with luminance → Typical Algorithmic Error (Commonly seen when incorrect curves are applied in RGB space instead of luminance space)

4. Category 2: Pure Color Channel Integrity Test

2️⃣ Input: Pure Colors and Single Channel Gradients

Construct:

(R=1, G=0, B=0)
(0,1,0)
(0,0,1)
And gradients for each channel 0→1

Expected Behavior

Film allows:
- Pure Red → Shifts to Orange / Magenta (This is style)
But it must be explainable and symmetric

Calculable Metrics

Metric C: Channel Leakage Matrix

You can approximate a 3×3 matrix:

[ R' ]   [ a b c ] [ R ]
[ G' ] ≈ [ d e f ] [ G ]
[ B' ]   [ g h i ] [ B ]

Filmr Knowledge Base