Card condition from photos

Send a photo of a card's front, and ideally its back. You get measured centering and edge whitening, the damage a vision model can see, and the grade range that evidence allows, with what the photos can't show. The same feature is available as the GraphQL field assessCondition, the MCP tool assess_condition and the CLI command raritydb condition.

query($front: String!, $back: String) {
  assessCondition(front: $front, back: $back) {
    gradeRange { low high } confidence limitingFactors
    centeringLimits { grader grade note }
    front { centering { leftRight { display uncertainty } topBottom { display uncertainty } bordersMm }
            wear { location kind whitening severity } cardShortEdgePx glareFraction }
    back { centering { leftRight { display } topBottom { display } } wear { location severity } }
    defects { kind side location severity confidence }
    cannotAssess disclaimer
  }
}

raritydb condition front.jpg --back back.jpg

Images are base64, a data URL or an https:// URL (JPEG, PNG or WebP, at most 10 MB). Each call counts as one request.

This is an estimate, not an official grade. raritydb is not affiliated with PSA, BGS or CGC; their names only describe the scale.

Taking the photos

  • Take the card out of its toploader or slab. A penny sleeve adds glare and hides edges.
  • Lay it flat on a plain background that contrasts with the border (dark for yellow or silver borders).
  • Use soft, even light from above, with no glare stripe across the card.
  • Fill most of the frame. The card must be at least 1000 px wide in the photo; 1500 px or more gives tighter centering.
  • Photograph the back too. Without it, back centering, edges and surface aren't assessed, and the range is wider.

What's measured

  • Centering: the card's cut edge and the inner edge of its border are found along 40 lines per side. Border widths give left/right and top/bottom ratios, e.g. 57/43 ±1.4. The ± comes from how consistent those lines are and the photo's resolution.
    • Full-art, borderless and textured-border cards have no clean border to measure. Those axes come back null and are listed in cannotAssess.
    • Consistency check: a Pokémon card's border is the same width all round, and a miscut only moves it from one side to the other. If left+right and top+bottom disagree by more than 25%, the inner edge was found inside the design (for example, a pale name bar next to the border). That axis is reported as not measurable rather than guessed.
  • Centering limits: the best grade each grader's published centering limits allow for the worse axis.
    • PSA: 10 needs 55/45 front and 75/25 back; 9 needs 60/40; 7, 75/25.
    • CGC: Pristine 10 needs 50/50; Gem Mint 10, 55/45 front and 75/25 back; 9, 60/40 front and 90/10 back; 8, 65/35.
    • BGS: publishes only the Pristine 10 limit (50/50 front, 55/45 back).
    • Grades without a published number are ruled out only by the next looser published limit.
  • Whitening: for each edge (a 1.5 mm band) and corner (up to a 4 mm square, kept within the border), the share of worn-white pixels against a coloured border. Glare blobs and anything matching the table outside the card don't count. On a white table, or with a white or silver border, whitening can't be told apart and is reported not_measurable, so use a dark background. Severity is none, light (under 2%), moderate (under 6%) or heavy. White or silver borders are not_measurable.

What the vision model reports

Creases, dents, indentations, scratches, print lines, stains, writing, edge chips and corner wear, each with a side, location, severity and confidence. Only defects at least 50% likely are returned. It also lists what the photos can't show, e.g. "fine scratches need raking light".

How the range is set

  • Each factor caps the grade:
    • Centering: PSA's published limits.
    • Measured whitening: light 9, moderate 8, heavy 6.
    • Defects: for example, a minor crease caps at 6 and a dent at 7.
  • high: the best case the evidence allows. Centering is read at the favourable end of its ±, and only defects at least 80% likely count.
  • low: the worst case it doesn't rule out. Centering is read at the unfavourable end, defects at least 50% likely count, and it goes lower for what couldn't be seen:
    • one grade for a missing back
    • one for unseen surface or listed limits
    • one for unmeasurable centering
    • half a grade for heavy glare
  • confidence: high when the range spans one grade or less, medium for two, low beyond that. limitingFactors lists what holds the grade down.

Limits

  • Photos can't show fine surface scratches, slight dents without raking light, or anything hidden by glare or a sleeve. Graders also judge eye appeal.
  • There's no public set of graded-card photos to calibrate against, so treat the range as a pre-screen before paying for grading, not a promise.

This page as Markdown: /docs/condition.md