About Aiyi & Color Season AI
The Story Behind Color Season AI
Hi — I'm Aiyi, the person behind Color Season AI. I started it with one question: why does professional personal color analysis cost hundreds of dollars when the logic behind it can be worked out from a single photo in seconds?
Four Tries at Getting the Accuracy Right
I started building this in January 2026. Nine months and four vision engines later, here's the version you're using — the two I had to throw out, and the one I outgrew.
Google's open-source vision model
Jan → Mid-April 2026 · AbandonedStarted building in January 2026, planning to wire up Google's open-source vision model. Registered the colorseasonai.com domain in early April and pushed the first public version live. Ran it for about two weeks. The upside was real — lightweight and fast, easy on the user side. But it couldn't tell 12 sub-seasons apart cleanly, and the precision wasn't where I'd put my own name behind it. Shelved.
3 months of build · ~2 weeks live
A custom classifier I trained myself
Mid-April → End of April 2026 · AbandonedMid-April I swapped in a small classifier I trained myself. Rented a GPU, collected 5,000 sample selfies, did my own labeling, and ran the new model for about two weeks. It taught me what real data work looks like — and that 5,000 images plus my own labeling weren't nearly enough to generalize cleanly across 12 seasons. Shelved again.
5,000 selfies · GPU rental · ~2 weeks live
Measuring first, then letting a vision model judge
May → September 2026 · ReplacedIn May I stopped trying to do it myself. Google's MediaPipe started sampling your actual skin tone in the browser — cheeks, forehead, chin — and Meta's Llama 3.2 Vision on Cloudflare Workers AI read your face. A week to wire up, and accuracy jumped roughly 20 points, to about 80%. The summer taught me two more things. In August I found that asking the model to run the whole season decision made it ignore the photo, so I cut its job down to describing what it sees on four scales and moved the 12-season decision into code. In September I ran the same photo twice and got two different seasons, so every photo started being read three times, keeping the middle value of each reading.
Llama 3.2 Vision · ~80% accuracy · retired October 1
A bigger model that reads finer and holds steady
October 2026 → now · Live, still tuningOn October 1 I moved the reading to Google's Gemma 4 (26B), on the same Cloudflare platform. It costs more per photo than the model it replaced, and it earns it. On my 20-photo test set it used 13 different score combinations where the old model managed 9, and its warm-versus-cool reading tracked the measured skin color far more closely (a correlation of 0.83, up from 0.50). Ask it about the same photo again and it answers the same way — identical numbers on 17 of 18 repeats. So a photo is now read once, and only when that read lands close to the line between two seasons does it get two more reads, with the middle value deciding. Before any change to the model or the prompt goes live, a set of licensed sample photos runs through the real model and has to land where it should, and every live analysis logs how many reads it took and how far they disagreed. This is the version you're using right now.
A larger model · costs more per photo · re-reads borderline photos
What's Changed Recently
The analysis and the handbook both change every few weeks. Here is what moved, newest first.
Analysis
The reading moved to Gemma 4, a larger model that costs more per photo. Clear-cut photos are read once; borderline ones get two more reads and the middle value decides.
Handbook
The preview opens like a document before you buy: the free pages in full, then one locked page, so you can see what the rest of your book looks like.
Results
One photo, one season. The result page shows the season your photo reads as, instead of asking you to pick between close options.
Handbook
The handbook became a 26-page A4 book built from your own photo, colors and numbers: makeup and hair pages drawn from real images, garments drawn in your colors, and chapters six to ten written from your readings rather than a season template.
Handbook
You no longer need an account to buy the handbook.
Analysis
The same photo stopped coming back as two different seasons: randomness switched off, and each photo read three times with the middle value of every reading kept.
Analysis
The vision model now only describes what it sees on four 0–100 scales — warmth, depth, clarity and contrast — and the 12-season decision is made in code, against fixed thresholds that are tested on every change.
Why I Built This
Warm vs. Cool
Light vs. Dark
Soft vs. Clear
What the AI reads from your photo
There's a particular kind of morning I think a lot of us know: standing in front of a full closet, running late, and somehow having nothing to wear. For years that was me. I'd pull a sweater off the hanger, hold it up, put it back — not because it was a bad sweater, but because something about it just didn't sit right against my face, and I couldn't have told you why.
When I finally learned a little about personal color analysis — which undertones, depths, and shades actually suit me — getting dressed got quieter. I stopped buying things that would sit unworn. I started reaching for the same few colors and feeling, honestly, more like myself. It was never about owning more clothes; it was about understanding the ones I already loved.
I'm not a stylist or an image consultant — I'm Aiyi, someone who got tired of guessing, looked into how color analysis actually works, and realized the core of it — warm vs. cool, light vs. dark, soft vs. clear — is something a good vision model can read from a single photo. So I built the tool I wished I'd had: free, instant, no appointment, no sign-up.
If you've ever stood in front of your closet feeling that small, familiar frustration, I hope this gives you somewhere to start — and a gentler relationship with getting dressed.
The Cost of Traditional Color Analysis
Personal color analysis has existed since the 1980s, when color consultants began categorizing people into four seasonal types — Spring, Summer, Autumn, and Winter — based on their natural coloring. Over the decades, the system evolved into the more precise 12-season model that professional image consultants use today.
The problem? A single in-person session costs anywhere from $150 to $450, and availability is limited to major cities. Many people are curious about their season but can't justify the cost or travel. Online quizzes help, but they rely on self-reported answers (“Are your veins blue or green?”) that most people struggle to judge accurately.
That's the gap I built Color Season AI to fill. It's free, no account required, and your photo is analyzed securely. A signed-out photo is not saved; signing in lets you keep a reduced-size copy in your Style Profile.
How the AI Analysis Works
Upload a Selfie
Take or upload a photo in natural daylight, without heavy makeup or filters.
AI Image Analysis
A vision model evaluates visible undertone, depth, clarity, eye color, hair color, and feature contrast from the uploaded image.
Undertone Detection
The AI estimates whether your coloring reads warm, cool, or neutral while accounting for lighting quality and image conditions.
Season Classification
Combining undertone warmth, depth (light vs. dark), and chroma (muted vs. clear), the system maps you to one of 12 seasonal sub-types.
Your Palette
You receive a personalized color palette with your best colors for clothing, makeup, hair, and accessories — plus colors to avoid.
Want the full pipeline — how undertone, depth, and chroma are read, plus the real accuracy numbers and limits? See how our AI color analysis works.
What I Won't Do
I Don't Sell, Share, or Train On Your Data
Your analysis result is saved so you can find it again later — but it's not sold, shared with advertisers, or used to train any model, mine or anyone else's. Inference runs through Cloudflare Workers AI, which doesn't retain inputs for training. If you're signed in, the small copy of your photo kept for your Style Profile sits in private storage that only your account can open.
Your Photo Stays Only If You Sign In
Signed out, your selfie is read and then discarded — the photo is never saved, and the season and palette stay for 60 minutes. Signed in, a small 600px copy goes to your Style Profile so you can compare seasons on your own face later, and you can delete it from your result page at any time.
I'll Tell You What AI Can't Do
AI color analysis isn't infallible. Bad lighting, heavy filters, and unusual hair color throw it off. I'd rather flag the limits up front than sell certainty I can't deliver — try the quiz alongside the AI scan if you want a second read.
The 12-Season Color System
The 12-season system refines the classic four seasons by adding three sub-types to each. For example, Winter splits into Dark Winter (deep, warm-leaning), True Winter (high contrast, neutral-cool), and Bright Winter (vivid, clear colors). This granularity means your palette recommendation is far more precise than a simple “you're a Winter.”
Each sub-season is defined by three axes: undertone (warm vs. cool), depth (light vs. dark), and chroma (muted vs. clear). The algorithm I built measures all three from your selfie to place you on the seasonal spectrum with confidence.
Standing in front of your closet right now?
The 30-second scan is what I would have wanted on those mornings. Try it — or if you'd rather not upload a photo, the quiz works too.
Questions or feedback? Reach me here. Color Season AI is also listed on There's An AI For That, where you can see the tool's profile and leave a review.