In the previous piece, we got as far as "words alone won't do it." Which raises a natural follow-up: what if we just made the vocabulary more precise?
Over fit, semi-over fit, loose fit, relaxed fit, wide fit… Slice it finely enough, and surely one day it will cover everything?
So we actually counted. It doesn't. And the way it falls short is interesting.
For reference, semi-over fit (세미오버핏) is a term 37 sellers use 69 times, and the image analysis model has never used it once. Half-balloon fit (하프벌룬핏) is even further out.
01 · Count
We pulled every expression of the form "___ fit" from 15,000+ product descriptions and counted them. We looked separately at the drafts written by the image analysis model and the final versions sellers reviewed and edited.
the model uses
sellers use
These counts only include terms that appear at least twice. If you count the one-offs too, it becomes 32 vs. 205, and even after stripping out typos and grammatical fragments, 161 remain.
Here they are in order of frequency. The number in parentheses is how many times each appears.
Loose fit(2,535) · Over fit(1,987) · Slim fit(1,279) · Wide fit(826) · Balloon fit(179) · Regular fit(178) · Straight-line fit (일자핏)(129) · Muscle fit(99) · Semi-over fit(69) · Drapey fit (떨어지는핏)(51) · Straight fit(49) · Oversize fit(46) · Crop fit(38) · Stingray fit (가오리핏)(37) · Baggy fit(36) · Box fit(36) · Flare fit(33) · Basic fit (기본핏)(24) · Tight fit(22) · Pants fit (바지핏)(22) · Bermuda fit(20) · Proper fit (정핏)(19) · Bootcut fit(18) · Sleeve fit (소매핏)(17) · Boxy fit(17) · Jogger fit(14) · Semi-loose fit(10) · Semi-crop fit(10) · Semi-wide fit(10) …
The first four make up most of the total. But things get interesting from number five onward: balloon, stingray, baggy, flare, bootcut. The more distinctive the garment's shape, the further back its term sits. These words don't come up often, but when they do, they are the very identity of that garment.
The mix also includes words like proper fit (정핏), basic fit (기본핏), pants fit (바지핏), and sleeve fit (소매핏). They aren't in any dictionary; they're simply what sellers happen to call things.
What stands out even more is this: 84.9% of the terms sellers use are words the model never uses. The overlap is far smaller than you'd expect.
02 · The Tail
Lots of variety wouldn't matter if knowing a few common terms covered most cases. So we checked that, too.
| To cover this share of all uses | Model | Sellers |
|---|---|---|
| 50% | 2 terms | 2 terms |
| 80% | 3 terms | 4 terms |
| 90% | 4 terms | 12 terms |
| 95% | 5 terms | 27 terms |
| 99% | 11 terms | 123 terms |
Up to the halfway mark, the two are identical. For both, loose fit and over fit are all you need. But the higher you go, the further they drift apart. For the model, five terms get you to 95%. For sellers, it takes twenty-seven.
At 99%, it's eleven vs. one hundred twenty-three. However large you make the dictionary, something new keeps showing up at the end.
03 · Everyone Their Own
So who uses the terms in that tail? For each term, we counted how many sellers use it.
| Sellers using the term | Model | Sellers |
|---|---|---|
| Just 1 | 46.9% | 62.4% (128 terms) |
| 2-4 | 9.4% | 19.0% |
| 5-19 | 15.6% | 12.2% |
| 20 or more | 28.1% | 6.3% (13 terms) |
62.4% of the terms sellers use are used by that one seller alone. And only thirteen terms are shared by more than twenty sellers.
The model is the exact opposite. Since it's a single model writing everything, its terms are naturally shared, and 28.1% of them show up across twenty or more sellers.
The shared vocabulary comes to about thirteen terms; the rest are words each seller made up. A dictionary could hold those thirteen, but it can't hold the rest, because they haven't been coined yet.
04 · How They Make Them
New terms don't just appear at random. There's a pattern. Sellers take an existing term and put a word signaling intensity in front of it.
Semi-over fit (69 uses · 37 sellers) · Semi-loose fit (10·5) · Semi-crop fit (10·6) · Semi-wide fit (10·6) · Semi fit (3·1) · Semi-muscle fit (3·1) · Semi-regular fit (2·2) · Semi-balloon fit (2·2) · Semi-slim fit (2·2) · Semi-baggy fit (2·1)
The model has never used a single one of these ten. Yet 53 sellers use them. Semi-over fit alone is used by 37 sellers and appears 69 times.
Sellers also build terms in another direction.
Half-balloon fit (하프벌룬핏): balloon fit, but not quite that far
Loose-over fit · Big-size-over fit · Super-over fit: even more "over" than over fit
Loose-stingray fit · Wide-curved-jeans fit · Big-wide fit: two shapes layered together
When an existing term falls short, they put a word in front of it; when that still falls short, they stack another word on top of the one they added. Over fit → big-size-over fit: the intensity keeps climbing.
Half-balloon fit in particular picks up right where the previous piece left off. Even though the word "balloon fit" already existed, sellers needed a separate word for "not quite that, but something like it." That's how much the difference in shape matters to sellers.
This isn't one or two people's habit. Fifty-three sellers who don't know each other each arrived at the same way of building words. They needed a word for "over fit, but not that far," and all of them came up with it the same way.
And it isn't just about "semi-." We counted all eight intensity prefixes.
| Prefix | Terms sellers built | Terms the model uses |
|---|---|---|
| Semi- | 10 terms · 113 uses · 53 sellers | 0 |
| Big- | 4 terms | 0 |
| Super- · Half- · Crop- | 4 terms | 0 |
| Loose- · Wide- · Tight- | Several | 0 |
| Over- | 3 terms · 48 uses | 1 |
For seven of the eight, the model doesn't use a single term. "Over-" is the only exception, with one term (oversize fit). This kind of word-building happens only on the seller side.
Vocabulary is not a fixed list; it keeps being produced. However big you make the dictionary, sellers coin words faster than the dictionary can be built.
05 · What They Fix
We also looked at which fields sellers edit most often when reviewing the model's drafts.
| Field | Share edited |
|---|---|
| Fit & silhouette | 68.0% |
| Fit | 58.3% |
| Top design | 51.4% |
| Details | 41.5% |
| Material & color | 40.7% |
| Color | 23.2% |
Fit-related fields are at the top, and color is at the bottom. The model can tell color by looking, but not fit. It's the same story we saw in the previous piece: it gets right what sits inside the average, and misses what sits outside it.
And looking only at descriptions that contain a fit term, sellers change that term in 46.3% of them. Roughly one in every two.
Wrap-up
The numbers show why the approach of "just define the fit vocabulary well" doesn't work.
- Seller vocabulary is more than 5 times the model's, and 84.9% of it consists of words the model doesn't use
- The tail is long: you need twenty-seven terms to reach 95%, and one hundred twenty-three to reach 99%
- 62.4% are words only one seller uses, and the shared vocabulary is just thirteen terms
- New terms keep being made: of the eight intensity prefixes, the model never uses seven
Vocabulary is not something you can catch up with. The moment you try to finish the dictionary, new terms have already appeared. And since most of those terms belong to a single seller, putting them in a shared dictionary doesn't even make sense to begin with.
That leaves one clear direction: instead of chasing vocabulary, learn what each seller means by their words, from that seller's own clothes. It lands in the same place as the previous piece's conclusion: we need something that exists separately for each brand.
And the very fact that these terms keep being coined is raw material on our side, too. A new word means a new garment has arrived, and that's a garment the giant models don't know yet.