1. Trichromatic theory

Section 4 uses the CIE 1931 2° observer, tabulated at 5 nm with linear interpolation, and energy-weighted XYZ sums. Shared photon shares are converted to relative radiant power using 1/wavelength. This is a standard-observer chromaticity diagram, independent of the selected animal, cone blocks, intensity and adaptation. It is not a projection of Sections 2 or 3. The same nonnegative photon shares used by the receptor lesson drive this mixture. The source boundary is their convex hull in xy; with three non-collinear primaries it is a triangle. Relative light shares move the mixture inside that hull, while the main intensity slider scales total light without changing chromaticity. There is no separate target or matching task. The iPhone 17 overlay is the nominal Display P3 triangle, based on Apple's Wide color (P3) specification, not a measured device profile. Adjustable narrowband primaries illustrate alternative triangles. No fixed trio spans the entire spectral locus. Diagram fills use clipped sRGB approximations.

Section 2 plots the same dimensionless cone responses as its opponent meters: A=L−M, B=S−(L+M)/2, P=(L+M)/2. Inputs are the slider’s total photon intensity times each sample’s share, summed through the CIE cone sensitivities before R=D/(D+K). K follows the existing global light-level adaptation state. No per-spectrum peak normalization, exposure rescaling, local chromatic adaptation, or afterimage offset is applied to the coordinates. The reference contains all nonnegative mixtures of the tabulated 390–700 nm spectra over all intensities, transformed through the current K values. Thus it is an adapting response region, not a fixed reflectance solid or a calibrated perceptual gamut. For each P, linear spectral-cone inequalities constrain absorbed L,M,S drives; inverse response odds determine the lower and upper S response for each sampled L/M balance. The outline samples this nonlinear boundary; membership tests use the original inequalities. The P=1 line denotes the infinite-drive closure, not finite physiological output. Source dots represent each source alone at its allocated photon share; they have their own P and are projected in slice mode. Forced cone inputs can lie outside the ordinary-spectral region. Reference colors invert the responses back to linear drive for an approximate CIE/sRGB hue cue; colors are not inferred from equal responses or the opponent signs. The external appearance swatch retains its separate illustrative low-light rendering. These response axes do not predict perceived hue or brightness.

Section 3 uses an illustrative opponent-coordinate space, not CIELAB, CIECAM16, or a calibrated neural color-appearance model. White-relative linear cone inputs q=(qL,qM,qS) map to A=2(qL−qM), B=qS−(qL+qM)/2, and J=(qL+qM)/2. J is a brightness proxy, not luminance or a measured perceptual scale. Negative A or B means an opponent direction, not negative cone excitation. The neutral black point is (0,0,0). Equal white-relative inputs give the neutral axis. The finite reference volume is an approximate object-color solid: 0≤reflectance≤1 under equal-energy illumination, sampled at 1 nm over the embedded CIE 390–700 nm range. Cone inputs are divided by their perfect-reflector values. At each J, a fractional-knapsack support calculation maximizes each chromatic direction subject to that J; 64 support directions form the displayed slice polygon. Stacked slices approximate the 3D surface. This is a defined surface-color reference, NOT the complete gamut of ordinary emitted lights, all adaptation states, or all human perceptions. Narrowband light sources may lie outside it without being unusual afterimages. Object-color solids and optimal reflectances. The ideal M-only toggle indicates a direction related to olo, not a measured olo coordinate or an RGB rendering of olo. Cone-targeted stimulation.

The Magenta and Purple presets mix 450 and 650 nm with different photon proportions; neither example is a single-wavelength stimulus (spectral violet is different from this red–blue purple). The metamer recipes match the three linear human cone drives and total photon intensity in this CIE 2° reference: A uses 480 and 590 nm, B uses 450, 530, and 650 nm. Their LMS match is numerical to floating-point precision, not a measured match for every observer. The target swatches share the same approximate screen conversion as the mixture preview. The buttons select bright human vision and clear artificial input blocks. Rod excitations need not match, so this is a photopic example, not a scotopic match. Samples added by hand default to equal relative weight; preset proportions remain attached to their widgets when wavelengths move. Light-share controls set photon percentages, redistributing the other samples proportionally while keeping total stimulus intensity fixed. Zero-share samples remain editable; a single remaining sample receives 100%. These controls do not modify the receptor curves, which remain full-intensity single-wavelength tests.

Churchland, P. (2005). Chimerical colors: Some phenomenological predictions from cognitive neuroscience. Philosophical Psychology, 18(5), 527–560. Conceptual reference for the chimerical-afterimage demonstrations and opponent-space interpretation; it does not validate this visualization’s equations, time constants, or plotted coordinates.

The 30-second exercise displays an actual RGB inducer followed by its physical probe. It never paints the predicted afterimage onto the probe. Screen RGB is decoded and converted to LMS, then divided by nominal screen-white LMS before entering the same A,B,J transform as the reference volume. Comparing these white-relative inputs with the equal-energy reference assumes diagonal cone-wise white normalization; this is not a measured cross-illuminant appearance match. Existing cone gains g approach 2/(1+D/Dgray), with 10-second time constants; the letters, receptor probes, and cone meters continue to show that first-stage model. The appearance plot adds a separate illustrative postreceptor memory h: dh/dt=(0.25·[A(q),B(q)]−h)/10 seconds. Its point is [A(gq)−hA, B(gq)−hB, J(gq)]. Section 2 is a separate response-space view described below; it does not use the local chromatic gains or opponent memory from this exercise. Memory is integrated analytically across the inducer/probe transition and reset with the model control. The coefficient, kinetics, gain rule, and brightness proxy are teaching assumptions, not fitted physiological parameters. This offset produces a decaying chromatic residual at J=0 on an ideal black probe; no negative photons or negative S-cone excitations are needed. Blueward displacement at J=0 illustrates a hypothesis about black-level afterimages; it does not measure “darker than black.” Modern afterimage studies favor cone adaptation for tested gray-probe hue matches, so this combined model must not be presented as an established quantitative account of these unusual appearances. Afterimage mechanisms and cone-adaptation evidence. Reference slices and points share the same coordinate definitions; only the point carries the current adaptation state. An adapted point outside the fixed reference is a model result, not proof of a novel or impossible percept. Slice mode follows the point’s J; 3D mode preserves the chosen viewing angle during animation. Unadapted reference markers at another J are explicitly identified as projections in slice mode. Afterimage coordinates are not given invented RGB colors.

The human dichromacy options omit a functional cone contribution: protanopia omits L, deuteranopia M, and tritanopia S. Remaining human cones retain the same CIE fundamentals, light adaptation and nonlinear response as the trichromatic model. The comparison is S−M, S−L or L−M respectively; strength is M, L or mean L/M. These are teaching coordinates, not fitted neural circuitry or clinical luminance functions. One normalized ratio describes the remaining color balance. The strip and circular cue use a schematic palette, not a calibrated dichromat appearance simulation; a missing cone is not a missing RGB primary. Rods remain available. Scene panels and wavelength colors remain trichromatic human references; the separate trichromatic afterimage lesson is hidden in two-cone modes. Anomalous trichromacy and individual variation are not modeled. Brettel, Viénot & Mollon (1997) describe a calibrated approach to appearance simulation, distinct from this response demonstration.

The Deer option is a simplified white-tailed deer cone model based on Jacobs et al. (1994): approximate peaks of 497 nm for rods, 455 nm for the short-wave cone (within the reported 450–460 nm range), and 537 nm for the middle-wave cone. It removes the separate third cone from the graphs and signal demo rather than treating its missing response as an ordinary zero input to human opponent circuitry. Its color comparison is the illustrative S − M and its strength proxy is M. With two cone classes, normalized color balance has one independent coordinate, so the human 3D view is replaced by a one-dimensional strip. Strip colors are a human-readable encoding, not evidence that deer experience those human hues. Gaussian widths, saturation parameters, adaptation kinetics, and the M-only strength weighting are teaching assumptions, not fitted deer physiology or measured deer luminance. Switching cone model starts receptor gain at equilibrium for the current adapting light and resets traces. The scene panels, foveal anatomy, and CIE spectrum/luminance references remain human illustrations. Jacobs et al.: deer spectral mechanisms (DOI: 10.1007/BF00217375).

The photoreceptor graphs and opponent meters retain a dimensionless adapting response model R=D/(D+K). The three comparisons L−M, S−(L+M)/2, and (L+M)/2 are illustrative graded-response combinations, not calibrated perceptual red–green/blue–yellow axes or firing rates. To calculate color, the code reconstructs response L=P+A/2, M=P−A/2, S=P+B, then recovers linear drive D=KR/(1−R) using the current model gains. This inverse is a mathematical conversion for the visualization, not a claim that the brain literally performs these calculations. Linear human LMS drives use the CIE cone-based 2° reference, reconstructed at 1 nm from the official CIE 2015 XYZ table and verified against the CIE 2006 5 nm LMS table (maximum component difference below 3×10⁻⁵), converted from energy to photon weighting with the common 550/λ factor. A cone-fundamental-based LMS→XYZ transformation then gives the CIE 2015-style reference, followed by nominal sRGB conversion. The forward transform reproduces the official 1 nm XYZ table within numerical roundoff. Spectral marker colors and the human swatch use this single transformation. Section 3 reference colors are approximate screen cues for white-relative coordinates. Section 2 reference colors recover linear drives from the adapting responses. Neither plot’s axes are calibrated hue labels. Equal L and M is not defined as unique yellow; equal LMS is not a physical white reference. The mean L/M response bar is a teaching quantity, not CIE luminance. Section 2 bar lengths encode signal magnitude; fixed red/green and blue/yellow endpoint hues identify channel direction, and the mean-response bar is neutral. Their display RGB values share the swatch’s adapted brightness multiplier, independently of bar length. This is a UI convention, not a separate color-appearance prediction for each channel; the bars retain their identifying hues rather than adopting the swatch’s rod-weighted desaturation. The swatch retains the separate illustrative illumination/adaptation rendering and rod-weighted achromatic blend; brightness, cone-response kinetics, mesopic hue interactions, and deer appearance are not calibrated predictions. CIE cone fundamentals; matching CIE cone-based XYZ data.

Photoreceptors release chemical messenger in the dark; light reduces this output. Each icon follows its card's current response proxy, including wavelength selection and gain adaptation. Dot visibility is proportional to 1 − response. This inverse relationship is illustrative, not a fitted synaptic release curve; dot movement is not a physical release rate. Reduced-motion settings stop dot travel. The connecting lines summarize shared retinal processing, not individual anatomical connections. Ganglion cells encode processed signals as electrical spikes carried along the optic nerve. The spike glyphs are schematic and do not simulate ganglion-cell firing rates or timing. Phototransduction; Retinal output.

The spectrum strip, wavelength-sample colors, human excitation diagram, and swatch reference hue share the CIE 2006 LMS / cone-based XYZ color pipeline. The matching CIE tables begin at 390 nm, so the stimulus widgets and plotted horizontal range are restricted to 390–700 nm rather than inventing cone sensitivities below 390 nm. Long-wave S values are zero where the published matching XYZ table has zero Z; this also agrees with the undefined long-wave S entries of the tabulated LMS reference. Nominal sRGB display conversion uses the ICC matrix and sRGB encoding; negative channels are clipped and the largest positive linear channel is scaled to 1. Many monochromatic colors are outside sRGB, and modern cone-based XYZ is not a literal measured-monitor characterization under the historical CIE 1931 observer, so this is explicitly an approximate screen rendering of the common colorimetric reference. It guarantees internal agreement for a single selected wavelength, not exact spectral appearance on an arbitrary display. The legacy CIE 1924 V(λ) view remains separately identified as standard photopic weighting. CIE LMS data (DOI:10.25039/CIE.DS.tijidesg); ICC display conversion.

Foveal view is a whole-image teaching model, active only while the pointer is inside the right scene and the checkbox is checked. The same rasterized scene, including dots, surface marks, shading, and silhouettes, passes through one spatial pipeline. Linear-light luminance is separated into four detail scales (approximately 1, 3, 8, and 20 SVG units). Smooth contrast thresholds suppress weak detail according to illumination and distance from simulated fixation. A small central rod-sensitivity deficit raises dim-light thresholds; off-center rod support improves faint-detail visibility while retaining poorer fine resolution. Strong contrasts can survive where weak details disappear. Large-scale local shading is preserved rather than replaced with a black or gray disk. This coarse local reconstruction is only an approximation to missing-detail appearance, not a model of cortical filling-in. Chromatic differences are pooled over larger neighborhoods toward the periphery, retaining broad uniform colors rather than making the entire periphery gray. All pixel buffers are opaque and blur uses clamped edge samples, avoiding transparent SVG filter borders. The existing global adaptation, rod/cone color blend, exposure mapping, and pupil animation still determine the scene input; local spatial thresholds depend on the illumination regime and do not have separately fitted adaptation kinetics. The camera image and 3D plots are unchanged. The cursor is hidden, with thin coordinate guides and a gap at fixation. Leaving or disabling foveal view restores the reference scene. A 400 × 190 image is processed; results are cached and refreshed during adaptation. Image size, contrast thresholds, visual-field extent, and time constants are illustrative, not calibrated retinal angles, luminance, measured contrast-sensitivity functions, or detection probabilities. Eye tracking, crowding, neural noise, individual differences, and complete mesopic color interactions are omitted. Uniform-field preservation and selective loss of low-contrast detail are engineering constraints, not proof of perceptual accuracy. Resolution across the visual field in dim light; Peripheral color and stimulus size; Foveal rod scotoma and filling-in.

Luminance contribution uses the CIE 1924 photopic V(λ) function from CIE 018:2019, sampled every 5 nm from 380–700 nm and embedded in this file. CIE data and attribution (DOI: 10.25039/CIE.DS.dktna2s3). At each wavelength, the plotted relative contribution is P × V(λ), where P is relative radiant power for an otherwise identical monochromatic viewing geometry. The common 683 lm/W factor is omitted because the display is relative; these values are not calibrated cd/m². The 3D vertical scale is logarithmic to display the full intensity range. The 2D cross-section uses a linear vertical scale that follows intensity, with actual relative values; curves are not divided by their peaks.

The intensity convention is view-specific: the receptor model uses photon-based test intensity, while the luminance view compares equal radiant power. The same slider number is not a conversion between those units. A photon-based luminance calculation would include photon energy hc/λ. Luminance mode applies fixed photopic weighting throughout its intensity axis; it does not model scotopic/mesopic sensitivity, so viewing-regime bands are omitted in this mode. V(λ) is independent of the schematic adaptation animation, and is not obtained by summing our approximate cone curves.

The two panels share a gold pyramid and equally sized red and blue-green spheres under neutral lighting. The left panel illustrates a fixed-exposure camera using a fixed compressed display mapping, independent of the eye’s adaptation state. The slider’s 14 orders of magnitude are compressed into 2.5 display-exposure decades, with a smooth highlight shoulder and sRGB encoding to keep material differences visible through most of the range. A fixed smooth shadow fade over the first 35% of slider travel keeps the dim camera scene from being lifted by display compression; above that point the compressed middle-range mapping is unchanged. A smooth blend toward white over the final 18% of slider travel illustrates overexposure. This compression, shadow fade, and washout are teaching choices, not a literal camera sensor response or a calibrated prediction of clipping; sensor noise is omitted. Both scenes share a smooth display-only fade over the lowest 8% of the log-intensity slider and become black at its minimum; that minimum is still a positive model intensity, not physical zero. This fade is a presentation choice, not a calibrated detection threshold. This is a visual cue on the arbitrary intensity scale, not calibrated luminance or a second simulation of vision. The right panel blends the material color with a neutral rod-brightness cue as light falls. The spheres have similar illustrative photopic luminance before tone mapping (RGB-weighted values about 0.25 and 0.26), but deliberately different rod weights (0.06 and 0.60): the blue-green sphere becomes relatively brighter than the red sphere in dim vision. These hand-selected display material values illustrate the Purkinje shift; RGB does not specify a unique spectrum, and this is not a spectral or calibrated color-appearance calculation. The smooth cone weight shares the spatial-vision transition and depends only on the selected illumination regime. Receptor gain recovery is not used as a color-saturation estimate: its earlier use could produce an unsupported transient color-to-gray change. Brightness continues to evolve during adaptation through the separate exposure model; channel tone mapping can temporarily wash out colors after a large light increase. This scene does not simulate the time course of chromatic adaptation or mesopic rod–cone color interactions. Pure rod rendering is achromatic, not blue-tinted. This simplified blend omits measured mesopic hue interactions and cannot predict an individual observer's perception. Measured Purkinje shift; Rod contributions to mesopic color perception.

A real scene's luminance is calculated from its spectral radiance weighted by V(λ). Perceived brightness additionally depends on adaptation, surrounding contrast, and other neural processing. Our colored scene retains neutral broadband lighting and a separate illustrative appearance model: it is not a monochromatic stimulus, does not change spectrum when switching graph views, and is not numerically driven by the luminance surface.

The plots are a phenomenological teaching model, not measured human responses. R(λ,I,t) = I A(λ) / [I A(λ) + K(t)], with cone sensitivity A and an adapting half-response parameter K. Human cone spectral sensitivities use the CIE 2006 reference; the response nonlinearity and adaptation parameters remain teaching assumptions. Rod and deer sensitivities remain approximate Gaussians. Response is a dimensionless saturating electrical-response proxy, not firing rate, voltage, or a predicted incremental flash response on a background. The shared ceiling of 1 is a model convention; curves are not independently normalized.

The scotopic, mesopic, and photopic labels and softly blended range bands indicate approximate dim, intermediate, and bright viewing regimes. Their positions are layout choices on an arbitrary intensity scale, not measured transition points; real regimes depend on adaptation, spectrum, and viewing conditions. The slider sets a neutral broadband adapting light B and the selected test intensity I = B. The surface intensity axis explores hypothetical monochromatic test stimuli at the current adaptation state; changing positions on the surface does not independently adapt each point. For each receptor, background drive is B times its mean sensitivity over 390–700 nm (equal photon weighting). Human cones use measured standard cone fundamentals with photon-energy conversion; rods and the deer teaching mode retain Gaussian approximations. The model relaxes log10(K/K₀) toward log10(1 + background drive/K₀). Rod gain adjustment is capped at 100× to retain bright-light saturation; cone adjustment is uncapped over the displayed range. K₀ is 1 for rods and 1000 for cones in arbitrary units. These choices are illustrative, not fitted physiological constants.

Receptor sensitivity uses separate illustrative time constants: rods 5 s toward lower sensitivity and 120 s toward higher sensitivity; cones 1 s and 10 s, respectively. All three cone classes share kinetics because this demonstration does not model subtype differences. This represents gain control only: it omits detailed calcium/cGMP kinetics, pigment bleaching and regeneration, background-current suppression, and neural processing. It cannot predict full human dark adaptation. Research motivating the qualitative distinction: experimental cone adaptation and rod/cone differences.

Both panels show the same matte pyramid under a fixed light direction. Face shading, cast shadow, and faint surface marks represent reflected light. The left panel maps logarithmic light level through a fixed display-exposure curve. The right uses a separate illustrative exposure and tone-mapping model to show lost and recovering detail; it is not computed by adding receptor responses and is not a calibrated prediction of subjective brightness. The slider minimum is a display-black floor, not zero physical light. Both scenes are black at that floor. The eye icon is a UI indicator.

Perceptual adaptation retains the fast/slow time constants of 2/30 simulated seconds after brightening and 8/20–600 seconds after dimming. These are heuristic, separate from receptor gain control. The speed control advances the light-level models on one clock (default 5×, maximum 10×); the separate afterimage viewing exercise uses real seconds. The timer resets at each light change and stops when all perceptual, pupil, and receptor states reach their numerical settling tolerances (0.002 for panel/pupil states and 0.001 log10 units for receptor gain). Speed changes preserve history. Loading starts fully adapted to the selected intensity. Elapsed minutes are model time, not a clinical dark-adaptation measurement.

Research informing the qualitative behavior: brightness fading during adaptation, dark adaptation and lighting, and brightness and temporal context. The white page and your own adaptation affect how these panels appear.

The small eye is a schematic pupil indicator: brighter ambient light produces a smaller pupil; dimmer light produces a larger pupil. Its separate illustrative response uses time constants of 2 simulated seconds for constriction and 6 for dilation, on the same adjustable simulation clock. Pupil size is not a measured diameter, does not alter the plotted receptor responses, and is not a proxy for the slower retinal adaptation.