Van Gogh 1881–1890 — VTL Full Analysis

A perceptual and structural reading of Van Gogh

Russell Parrish · Artist Influencer · artistinfluencer.com · Images and perceptual data: Lev Manovich / Cultural Analytics Lab

The Premise

This artifact maps the complete recorded output of Vincent van Gogh, 776 paintings, nine years, five biographical periods, through two measurement systems simultaneously. The left panel shows brightness and saturation, the perceptual surface of each painting. The right panel shows centroid offset and cohesion, the structural organization of visual mass computed from edge-gradient data.

Both coordinate systems are plotted at career scale. Period toggles and category filters allow any segment of the data to be isolated. A live statistics bar tracks mean cohesion per period as filters change, so structural differences can be verified across subject types. The question is not whether Van Gogh changed, the perceptual record makes that obvious, but whether the structural coordinates reveal a pattern the perceptual record cannot see.

The perceptual arc has a known shape. The structural arc has a different one. Both are present simultaneously in the same 776 paintings.

Five Periods

Early 1881–85	μ̄ = 0.411	Netherlands. Dark palette, peasant subjects, figure studies. 196 paintings.
Paris 1886–88	μ̄ = 0.333	Impressionism exposure, Japanese prints, brightening palette. 199 paintings.
Arles 1888–89	μ̄ = 0.317	Southern light, wheat fields, sunflowers, the Yellow House. 161 paintings.
Saint-Rémy 1889–90	μ̄ = 0.210	Asylum. Cypress trees, swirling marks, the Starry Night. 138 paintings.
Auvers 1890	μ̄ = 0.177	Final 70 days. Stormy wheat fields, village scenes. 82 paintings.

The Structural Arc

Mean cohesion (μ̄) across all five periods, a monotonic descent with two steeper drops.

Early

0.411

→

Paris

0.333

→

Arles

0.317

→

Saint-R.

0.210

→

Auvers

0.177

Two steeper drops: Early → Paris (−0.078) and Arles → Saint-Rémy (−0.107). The Arles-to-Paris segment is nearly flat (−0.016), the slight recovery that portrait data makes visible. The career ends 57% below where it began, structurally.

Navigation & Controls

Scroll wheelZoom toward cursor, each panel independent

Click + dragPan, each panel independent

HoverImage preview, title, period, category, axis values

Period togglesShow or hide individual periods; stats update live

Category togglesFilter by subject type: landscape, portrait, etc.

Rescale axes"To selection" rescales both panels to visible data; "Fixed" holds the full-career range

Size sliderResizes thumbnails in both panels simultaneously

The Two Coordinate Systems

ImagePlot 2

Perceptual Space

Brightness × saturation. What the eye registers, the tonal and chromatic surface of each painting. Developed by Lev Manovich / Cultural Analytics Lab.

VTL Kernel

Structural Space

Centroid offset × cohesion. Where visual mass is placed and how unified it is, extracted from edge gradients before any subject label is assigned.

Metric Definitions

Metric	What it measures
Left Panel: ImagePlot 2
`brightness`	Median luminance (0–255). The perceptual arc is dramatic: Early ≈ 80, Arles ≈ 175. The chromatic awakening is the steepest gradient in the dataset.
`saturation`	Median color saturation (0–255). Arles peaks; Saint-Rémy pulls back. Brightness holds closer than saturation does. The structural descent continues regardless.
Right Panel: VTL Kernel
`Δx±`	Signed horizontal centroid offset (−1 to +1). Zero = mass centered. Early landscapes scatter widely; Arles and Saint-Rémy pull toward center as fields and flat subjects dominate.
`μ cohesion`	Largest connected structural region as fraction of total gradient mass (0–1). The primary metric. Descends monotonically across all five periods. The structural arc opposes the perceptual arc.
`SDI (tooltip)`	Spatial Dispersion Index: mean distance of mass from centroid. Key null result: SDI is essentially flat across all five periods. The mass does not spread, it fragments in place.

What the Metric Captures

At 200px, individual brushstrokes are gone. What cohesion measures is how gradient energy, light/dark contrast, resolves across the image as a whole: concentrated into a unified field or fragmented into disconnected patches. At this scale, compositional organization and surface texture regime are both present and cannot be fully separated. The metric reads their aggregate. This is not a limitation to be corrected; it is what the system measures: overall visual field coherence, which is the combined output of where mass was placed and how the surface was treated. The career-long monotonic descent, the per-category controls, and the copy inversion all operate on that aggregate.

What to Look For

Early (blue): high μ despite dark palette. The Dutch figure studies and peasant compositions produce surprisingly unified structural mass; subjects are dense, figure-to-ground relationships are tight.

Paris (teal): first cohesion drop. Contact with Impressionist fragmentation and Japanese graphic flatness begins to diffuse structural mass. The perceptual record brightens while μ slides.

Arles (amber): perceptual peak, structural plateau. The big color explosion is visible in the left panel. μ holds from Paris rather than recovering, but portrait cohesion recovers here (isolate Portrait to see it).

Saint-Rémy (red): the steeper structural drop. Brightness holds. Saturation retreats. μ falls to 0.210. The swirling mark-making distributes gradient mass into fragmented patches invisible to perceptual measurement.

Auvers (purple): career floor. The stormy wheat fields of the final 70 days score the lowest mean cohesion of the career: 0.177. The descent does not stop at Saint-Rémy.

Per-Category Arcs

Toggling individual categories while watching the live stats bar reveals whether the career-long structural descent holds within subject type, or whether mixing subject categories produces an artifact. Use "Rescale to selection" when isolating a single category to bring the distribution into focus.

Category	Early	Paris	Arles	Saint-R.	Auvers
landscape	0.371	0.291	0.289	0.166	0.158
portrait	0.501	0.334	0.490	0.391	0.278
floral	0.619	0.421	0.398	0.224	0.060
still life	0.444	0.388	0.361	0.270	0.201
copy	—	0.388	0.412	0.280	—
urban	0.341	0.299	0.271	—	—

Green = relative high; red = relative low within that category's arc. "—" = fewer than 3 paintings in that cell.

The Structural Finding

The perceptual arc of Van Gogh's career has a familiar shape: the Dutch darkness, the Paris brightening, the Arles explosion, the continued intensity of Saint-Rémy, the final scatter of Auvers. This is one of the most documented chromatic transformations in the history of painting, and the left panel confirms it.

The structural arc runs in the opposite direction. Cohesion (μ) descends monotonically from Early (0.411) through Auvers (0.177), a 57% drop across the career. This is not a Saint-Rémy rupture; it is a career-long structural dissolution that happens to accelerate at two moments of biographical inflection: the move from the Netherlands to Paris, and the move from Arles to the asylum. The perceptual transformation and the structural dissolution proceed simultaneously in opposite directions.

Three specific results deserve attention. Floral is the most extreme trajectory: Early floral works score 0.619, the highest cohesion in the dataset, as tightly packed flower arrangements produce maximally unified gradient mass. Auvers floral scores 0.060, a 90% collapse, as loose cut-flower arrangements in the final period distribute mass almost randomly across the frame. Portrait shows an Arles recovery: cohesion drops from Early to Paris (0.501 → 0.334), then recovers in Arles (0.490), drops again in Saint-Rémy (0.391), and collapses in Auvers (0.278). The recovery in Arles is compositional: Van Gogh's Arles portrait practice produced unusually tight figure-to-ground relationships. Copy-as-scaffold confirms across the full dataset: the 27 Saint-Rémy copies after Millet, Delacroix, and others score 0.280 on average against 0.193 for concurrent own-work; the structural inheritance of the source composition is measurable.

The SDI (Spatial Dispersion Index) is flat across all five periods. The gradient mass does not spread as cohesion drops; it fragments in place. This rules out the most obvious alternative explanation: that the cohesion descent simply reflects compositions becoming more spatially distributed. It does not. The mass occupies the same average area; it breaks into smaller disconnected regions.

Perceptual brightness rises 118% from Early to Arles. Structural cohesion falls 57% from Early to Auvers. The two systems are measuring something genuinely different about the same paintings.

Coherence to Fragment