The Colour Myth

If you believe the styling internet, the single highest-leverage thing you can do to look better is fix your colours: get the palette right, match the undertone, wear the coordinates. It is a good story. This report tested it against the only thing that can adjudicate it — the actual numbers the model wrote on real outfits — and the story does not hold.

The reason is structural, and it is the most interesting thing in this report. OutfitScore does not score an outfit as one opaque number. It scores five named pillars, each out of 20, and adds them up. That means the data can answer a question almost no fashion dataset can: not "what did this outfit score?" but "which of the five things was actually holding it back?"

Section 1Methodology

The unit of analysis is one row in the OutfitScore analyses table. I keep a row if analysis_type is fashion or complete, analysis_result is present, and a numeric overall score can be read from overall_score, score, or total_score. Roast, makeup, body-type, and color-season analyses are excluded. Soft-deleted rows are not filtered: a deleted analysis is still a real observation of how the model labelled an upload.

The window is not a rolling 30-day panel. It is every qualifying row then in production: 27 September 2025 through 13 October 2026. That yields 21,775 scored rows before the pillar filter.

Score extraction follows the same numeric-regex guard used by seo_stats_service. Non-numeric strings such as "N/A" become null and drop the row. Each pillar score is read from analysis_result → detailed_breakdown (flat, the shape most fashion rows use) or analysis_result → outfit → detailed_breakdown (nested, the shape most complete rows use), resolving the key aliases the field has drifted through across scoring eras.

What "fully-scored" means, and why it matters

The model stores a pillar as locked — {"score": 0, "is_locked": true} — whenever it could not score that dimension on the photo. A lone garment, for instance, cannot be judged for styling cohesion or occasion. A stored 0 therefore means unscored, not zero. On most rows, one or more pillars are locked, so the breakdown does not reconcile with a comparable five-part total.

This report is restricted to rows where every pillar is present on the real 1–20 scale and the five pillars sum exactly to the headline score. That leaves 972 rows (4.5% of the 21,775 scored rows). This is the honest cost of doing the comparison properly: the pillars are only comparable to each other on outfits the model actually scored on all five. Those rows run 19 March 2026 to 10 October 2026.

The five pillars sum to the score

The mechanism is disclosed because it changes how to read every correlation in this report:

overall_score  =  fit_and_proportion
               +  color_harmony
               +  occasion_appropriateness
               +  styling_and_cohesion
               +  details_and_quality

# each pillar is 0-20; the five sum to a 0-100 overall score

Because the pillars compose the total, a pillar's correlation with the total is partly mechanical. That is exactly why this report leads with two measures that do not depend on the total: the separation between the score top-quartile and bottom-quartile, and how often each pillar is the weakest of the five on a given outfit.

Section 2The five pillars, side by side

Same 972 outfits. Same scale. Mean is out of a possible 20, so "share" is the mean divided by 20 — the fraction of the available points that pillar captured across the corpus. Q4−Q1 is the mean pillar score for outfits in the top quartile of overall score minus the bottom quartile: the cleanest answer to "how much does this pillar move with a good outfit?"

Two facts jump out. The two pillars people talk about least — styling cohesion and details quality — have the lowest means (12.84 and 12.74) and the highest correlations with the score (0.950 and 0.917). Colour harmony has one of the highest means (14.68) and a middling separation (6.04) — third of the five. The pillar the trend tells you to obsess over is the pillar outfits most often already pass.

Section 3The weak-link test

Separation tells you which pillar moves with a good outfit. It does not tell you where a typical outfit is actually losing points. For that, ask a blunter question of each row: of the five pillars, which scored lowest? That pillar is the outfit's weakest link — the one place the score left points on the table.

Ties are counted for every tied pillar, so the weak-link column sums above 100%. The pattern is unambiguous: details & quality is the weak link on more than half of all outfits (56.4%), and styling & cohesion on another third (35.2%). Colour harmony is the weak link on 9.6% — the least of any pillar — while being the strongest pillar on 29.7% of outfits, second only to occasion.

Colour is rarely the problem and often the strength. The points are lost on the two things nobody posts about: how finished the details are, and whether the pieces hang together.

Section 485-plus versus below 60

A third cut, to check the weak-link result a different way. Take the outfits that scored 85 or above (n = 143) and those that scored below 60 (n = 165), and compare the mean of each pillar in each band. The gap between them is how much that pillar differs between a strong outfit and a weak one.

Same story. Across the full 85-plus-to-below-60 span, colour harmony has the second-smallest gap (7.23), just above occasion (7.18) and below every execution pillar. Styling cohesion swings the most (8.71). Note the direction in the high band: at 85-plus colour harmony averages 18.75, the highest of any pillar. Strong outfits are strong on colour; weak outfits are not especially weak on it.

Section 5The colour distribution

Where do colour-harmony scores actually sit? The histogram is tightly clustered in its upper half. The modal score — the single most frequent value across 972 outfits — is 15 out of 20 (170 outfits), with 14 next (168) and 13 behind it (105). Roughly four in five outfits land at 13 or above. Scores of 10 or below are the minority: 79 of 972, about 8%.

This is what "colour is not the problem" looks like as a picture. A distribution that is right-skewed toward the top, with a long thin tail of genuinely bad palettes, is the signature of a dimension that people mostly get right. The dimensions that decide scores are the ones with fat left tails — the execution pillars, which sit at mean 12.74 and 12.84 against colour's 14.68.

Section 6What this report is not

This study has five close neighbours. None of them asks this question, which is why the page exists and why the framing is narrow.

The colour how-tos tell you which colours to wear. They cannot tell you whether colour is the thing costing you points, because none of them measured that. This page does not give palette advice and does not claim to. It reports what the five pillars did on the outfits scored on all five, and lets the reader draw the conclusion the trend pages are afraid of: your colours are probably fine.

Section 7Limitations

Coverage. Only 972 of 21,775 scored rows (4.5%) carry a reconciled five-pillar breakdown; the rest have at least one locked pillar. The findings describe those 972 rows and only those. They run from 19 March 2026 to 10 October 2026, so they lean recent. The locked-pillar majority is a different population — often lone garments or photos the model could not fully judge — and I make no claim that it behaves the same way.

The pillars compose the score. Because the overall score is the sum of the five, each pillar's correlation with the total is inflated by construction. That is why the headline evidence here is the quartile separation and the weak-link frequency, not the correlation. Even those are descriptive: they say which pillar moves with a good score, not that raising a pillar causes the score to rise. This is an observational snapshot, not a treatment study. No causal score lift is claimed anywhere on this page.

Locked is not zero. A locked pillar is stored as 0 but means "not scored." I exclude those rows rather than count them as zeros, which is the conservative choice: counting locked pillars as 0.0 would drag every mean down and make colour look artificially dominant, because colour is one of the pillars most often locked on a lone garment.

Model rubric, not human judges. The five pillar scores are written by a multimodal model against a fixed rubric, not by a panel with inter-rater reliability. The dimensions are the model's, and reasonable people could draw the boundaries between "cohesion" and "details" differently. The method's value is that it is consistent and that it is transparent about its five parts — not that it is the only defensible way to decompose style.

Defining "weakest" with ties. When two pillars tie for lowest, both count, so the weak-link shares sum above 100%. This slightly favours pillars that co-occur as joint-minimums. It does not change the ordering: details leads by a wide margin regardless.

Section 8Conclusion

On 972 fully-scored outfits, the overall mean is 69.5. Ranked by how far each pillar separates a top-quartile outfit from a bottom-quartile one, the order is styling & cohesion (7.07), details & quality (6.45), colour harmony (6.04), fit & proportion (5.92), occasion appropriateness (5.31). Ranked by how often each is the weakest link, details & quality (56.4%) and styling & cohesion (35.2%) lead; colour harmony is last (9.6%) and is the strongest pillar on 29.7% of outfits. Across the 85-plus-to-below-60 span, colour has the second-smallest gap of any pillar.

Every one of those cuts points the same way. The dimension the trend tells you to fix first is the one outfits most often already pass. The dimensions that decide the score are the unglamorous ones: whether the outfit hangs together, and whether the details look finished. That is not a reason to wear clashing colours. It is a reason to spend your next five minutes on a steamer and a mirror check instead of another palette quiz.