The TCGAI Score is an AI-generated estimate of a card's condition on a 10-point scale, with four sub-grades: centering, corners, edges and surface. It is not an official grade and it is not issued, endorsed or calibrated by any grading company. This page explains exactly how the score is produced, where the prices come from and where the method falls short, so you can decide how much weight to give it.
The short version
- You upload a photo of the front and, ideally, the back of the card.
- We check the photos, find the card and straighten it to its true proportions.
- Centering is measured in pixels. Corners, edges and surface are inspected by an AI vision model on zoomed crops, which lists every defect it finds.
- Fixed, published rules turn those measurements and defects into sub-grades. The TCGAI Score is the lowest sub-grade.
- The card is identified against public catalogues and matched with market prices.
1. Photo checks
Before anything is graded, the photos are checked for the basics: is there a card in the frame, is it sharp enough, is the exposure usable. Only impossible photos are rejected (no card, heavy blur, too dark or blown out). Smaller problems, such as glare or a slightly soft focus, do not stop the grading. They become warnings and widen the confidence range shown with the score.
2. Finding and straightening the card
The card's border is detected and the image is rectified to the standard 63 × 88 mm proportion, so a photo taken at a slight angle is measured as if it had been scanned flat. If the border cannot be found reliably, the result says so and the confidence range is wider.
3. Centering, measured in pixels
Centering is measured, not guessed: we compare the left-right and top-bottom borders on each face and take the worse of the two. The ratio maps to a centering sub-grade with separate tables for the front and the back, because graders are more lenient on the back. The tables follow the centering tolerances graders publish.
| Worst border ratio, up to | Front | Back |
|---|---|---|
| 55/45 | 10 | 10 |
| 57.5/42.5 | 9.5 | 10 |
| 60/40 | 9 | 10 |
| 62.5/37.5 | 8.5 | 10 |
| 65/35 | 8 | 10 |
| 67.5/32.5 | 7.5 | 10 |
| 75/25 | 7 | 10 |
| 80/20 | 6 | 9.5 |
| 85/15 | 5 | 9 |
| 90/10 | 3 | 9 |
| 95/5 | 2 | 6 |
| Worse | 2 | 4 |
On cards with no visible border (full art, borderless), pixel measurement is not possible. In that case the AI model estimates centering, and the result marks it as an estimate.
4. Corners, edges and surface: AI defect detection
The straightened card is cut into zoomed crops: the four corners and the four edges of each face, plus the full surface. A vision model (Google Gemini, run with its randomness set to zero so the same photo gives the same answer) inspects them and returns a structured list of defects. For each defect it reports the type (whitening, chipping, scratch, print line, dent, crease and so on), the face, the position, the severity and the approximate size. The same model also reads the printed card name, number and set.
Photos are used to produce your result and your certificate. See the Privacy Policy for how they are stored and who processes them.
5. From defects to a score: fixed rules
The AI does not choose the grade. Published, deterministic rules do, so two identical sets of findings always give the same score:
- Each defect costs points according to its severity (minor, moderate or major), its type and its size (larger than 3 mm and larger than 8 mm cost more).
- Within each category and face, the worst defect counts in full and the others count at half weight.
- Sub-grades are rounded down to the nearest half point.
- The TCGAI Score is the lowest sub-grade, because in practice one weak area caps a card. A secondary score (the average, never more than one point above the lowest) is shown for context.
- Structural damage caps the score no matter how clean the rest is: a bend caps it at 6, a crease at 4, water damage at 3, a tear at 2, and missing pieces or alterations at 1.
6. The confidence range
Every score comes with a range. It starts at ±0.5 and grows when centering had to be estimated, when the card outline was not detected, when the photos triggered two or more warnings, or when there is no back photo. It never exceeds ±2. A wide range means: take better photos before deciding.
7. Identity and prices
The card is matched against free public catalogues: TCGdex for Pokémon, Scryfall for Magic: The Gathering, YGOPRODeck for Yu-Gi-Oh!, Lorcast for Disney Lorcana and TCGplayer catalogue data for One Piece and Dragon Ball. Sports cards and anything unmatched are looked up on eBay.
- Raw price: from the catalogue's market data (TCGplayer, or Cardmarket converted to US dollars with the European Central Bank rate).
- Graded prices: from current eBay listings of graded copies. These are asking prices, not completed sales, so treat them as a reference range rather than a valuation. Outliers are removed before the range is calculated.
Prices and identities are cached for 24 hours.
8. Known limits
- The result can only be as good as the photos. Glare, low light, motion blur, a dirty lens or a sleeve all hide defects.
- A photo cannot show what a grader sees under magnification and angled light. Fine print lines and micro-scratches are the most common misses.
- The rules are not yet calibrated against a large set of officially graded cards. That work is ongoing, and every certificate records the version of the rules and of the AI instructions that produced it, so results can be traced when the method changes.
- Third-party grading companies use their own standards and human judgement. Their grade may be higher or lower than the TCGAI Score. See the AI estimate disclaimer.
Who runs TCGAI.PRO and how guides are written
TCGAI.PRO is an independent tool operated by Jose Camacho Sanchez in Spain (see the imprint). Guides on this site are written and maintained by the TCGAI.PRO team and signed as such. Each guide shows the date it was last updated. If you find an error in a guide or in a result, write to [email protected] and we will correct it.
Last reviewed: October 23, 2026.