ChatGPT color analysis.
Yes, ChatGPT can do a color analysis: upload a selfie, paste a prompt, and it names a season with a fluent explanation. The prompts are below, free. What it can't do is control your lighting — it reads the photo it's given, room and all — which is why the same face often collects different seasons on different runs. Use it as a first opinion; measure when you want the answer.
The prompts
No gate, no newsletter — here are prompts that get ChatGPT's best attempt. They work in any image-capable model (Gemini and Claude included). Ask for reasoning and uncertainty, not just a one-line answer: a model made to show its working produces a more useful read, and its hedges tell you where the photo is failing it.
Run a full 12-season personal color analysis on this photo (the sci\ART-lineage system: Light, True, Bright, Soft, and Dark variants of Spring, Summer, Autumn, and Winter). Work through five stages, in order, showing your reasoning at every stage. Do not skip ahead to the answer. STAGE 0 — AUDIT THE EVIDENCE. Before judging me, judge the photo. Name the likely light source (daylight / warm indoor bulb / mixed), say whether the white balance looks shifted and in which direction, and note anything that could be masking my natural coloring: makeup, filters, hair dye, a tan, heavy compression. Then grade the photo's data quality A to D and state what that grade caps your final confidence at. A single uncalibrated photo never earns more than “moderately confident,” no matter how clean it looks. STAGE 1 — RAW OBSERVATIONS ONLY. Describe what you literally see, before any interpretation: my skin's apparent temperature compared against the whites of my eyes and any white or neutral clothing in frame (never the background); my hair's depth and whether it warms or cools near the roots; my eye color and any visible pattern in it; and the natural spread of light-to-dark between skin, hair, and eyes. No season names allowed in this stage. STAGE 2 — THE FOUR AXES. Place me on each axis separately, each with its own confidence (low / medium / high) and the single strongest observation behind it: 1. Undertone (warm ↔ cool ↔ neutral) 2. Depth (light ↔ deep) — skin and hair taken together 3. Clarity (clear ↔ muted) — saturated and bright versus softened and grey-blended 4. Contrast (low ↔ high) — the spread between skin, hair, and eyes If an axis is a genuine coin flip from this photo, say “coin flip” — that is more useful than a fake lean. STAGE 3 — THE TOURNAMENT. Start with all twelve seasons on the table. Eliminate them in rounds, naming the axis that cuts each one, until three finalists remain. Then run paper drape rounds: for each pair of finalists, pick one color that belongs to one season's palette but would fail in the other's, and reason about which would sit better against the face in this photo. Keep a running shortlist with rough percentages — and keep them honest: they must reflect your stage-0 cap, not stack up to false certainty. STAGE 4 — THE READING CARD. Output, in this order: • Season and runner-up — they should usually be sister seasons (neighbours sharing an undertone); if yours aren't, explain why that happened. • The four axes, each with its call and a confidence word. • Evidence log: the three observations that did the most work. • Limits: three specific ways this photo could be fooling you, and which part of the answer each would flip. • Five colors from my likely season to try near my face, three to avoid. • The one question a second photo in different light would settle. Rules for the whole reading: never claim more certainty than the photo supports; if the lighting is warm or mixed, mark the undertone read provisional; and do not flatter me — a wrong season costs real money at the checkout.
Undertone: warm — moderately confident, driven by skin against collar contrast
Depth: medium-deep · Clarity: muted — lower confidence, see below
Likely season: True Autumn · runner-up: Dark Autumn
“The warm indoor lighting in this photo may be inflating the warmth I'm reading…”
Privacy is worth a beat here: a photo pasted into a chatbot is uploaded to its servers under that provider's terms. If that matters to you, read them first — or use tools that keep the photo on your device.
What it does well
Credit where due: the model knows the theory. It can recite the twelve seasons, reason about undertone and contrast, and explain why mustard fights a cool face better than most quiz sites can. For vocabulary, for understanding what a season even is, and for a fast first opinion at zero cost, it's genuinely useful — which is exactly why so many people try it before anything else.
Why the same photo collects different seasons
The failure mode isn't intelligence — it's input. A camera records the light in the room as much as the face in it, and an LLM reads the pixels it's handed with no way to know what the room did to them. Warm bulbs push skin golden; overcast windows push it cool; two photos of one face can honestly support two different seasons. Ask it twice and the sampling that makes chat feel natural adds its own wobble on top.
This is the same reason all selfie-based analysis disagrees with itself — the mechanism has a name, white balance, and it bends every uncalibrated read. The fix isn't a smarter reader; it's a controlled input: a known-neutral reference in the frame that lets the light be corrected before anything is measured.
ChatGPT vs a calibrated read
| ChatGPT with a prompt | A calibrated reading | |
|---|---|---|
| Lighting | Reads the photo as lit — the room comes along | Corrected against plain white paper in the frame |
| Repeatability | Can vary run to run on one photo | Same pipeline, same math, every time |
| Uncertainty | Only if you ask — and it's a self-report | Error bars and a confidence cap, printed |
| Cost | Free | Free instant read · $39 for the full bound reading |
ChatGPT is a well-read friend squinting at your photo; a calibrated read is a measurement. A friend's opinion is a fine place to start — keep the prompt, use it, and when the answers start disagreeing with each other, that's the moment the measurement earns its keep.
Does Gemini or Claude do color analysis?
Yes — any image-capable model will attempt it, and the prompts on this page work in all of them; the builder above has a toggle for each. The mechanism is identical, which means the limit is identical too: every one of them reads your photo exactly as lit, with no way to correct the room out of it. Switching models gets you a second opinion, not a calibrated one — useful for spotting agreement, not for settling a disagreement.
You have a season name. Now check it
Whatever named your season — ChatGPT, an app, a quiz, or a hunch — the name is only useful once it survives contact with real colors. Read your season's page and see whether its palette looks like your closet's best days; run a garment you already trust through the hex checker and see which decks actually claim it. If it holds up against both, wear it with confidence. If it doesn't, that disagreement is data — and a calibrated read is how you settle it.
Questions
People upload a selfie to ChatGPT with a prompt asking for their color season, and share the results — it spread through TikTok, Instagram, and style blogs as a free alternative to paid analysis. The appeal is real: instant, conversational, zero cost. The catch is the same one this whole page is about: the model reads the photo as lit, so results vary with the room and between runs. The prompts above are the honest version of the trend — they ask the model to show its reasoning and flag its own blind spots.
Sometimes — and inconsistently, which is the practical problem. The model reasons well about color theory, but it reads your photo exactly as lit, so warm bulbs, window light, and filters swing the result, and repeat runs can disagree on the same image. No published accuracy figure exists for LLM color analysis, so treat any confident-sounding answer as an estimate without an error bar.
One that asks for reasoning and uncertainty, not just a season. Have it estimate undertone, depth, and clarity separately, name a runner-up season, rate its confidence per axis, and flag anything about your photo's lighting that could be skewing the read. That last request matters most: it makes the model tell you when the input, not the analysis, is the weak link. A paste-ready version is on this page.
Two reasons stack. Your photos differ — each one carries its room's lighting, and white balance shifts apparent undertone — and the model itself samples its answers, so even one photo can collect different seasons on repeat runs. Disagreement between runs isn't evidence the whole idea is nonsense; it's evidence you've hit the limit of uncalibrated input.
Yes — image upload works on free tiers, subject to usage limits, and the prompts on this page are free to use anywhere. Callume's instant read is also free, no sign-up: the difference isn't the price, it's that a purpose-built pipeline measures in CIELAB and reports honest confidence rather than conversational certainty.
Maybe — the useful move is a second opinion under better-controlled input. If two uncalibrated tools agree, that's a real signal. If they disagree, the tiebreaker shouldn't be a third guess at the same broken input; it should be a read that corrects your lighting first and shows its confidence, so you can see how close the call actually was.