partforge

Photo cluttered real

hard tier · create · with a reference image

The request

In the attached photo the part I want is the black tray; the round grey knob beside it is something else lying on the desk — ignore it. The tray is 174.34 mm long, 107 mm wide and 17.98 mm tall: a plain open tray with 2.63 mm walls and floor and rounded corners, without the ribs or the magnets. Build it open side up, floor on Z = 0, length along X, centred on the origin. Name the sub-part `part`.
The photograph the model was given
Photo · made with iPhone camera · CC0-1.0

What the AI judge checks

  • The overall form matches the reference: the same kind of object, the same silhouette from the matching angle
  • Proportions match: the relative sizes of the major dimensions agree with the reference within what the eye can judge
  • Every major feature visible in the reference (holes, slots, bosses, tabs, lips) is present in the result
  • Those features sit where the reference has them
  • Nothing is present that the reference does not show — no decorative or unexplained geometry
  • The result is the tray the prompt named, not the round knob lying beside it, and the knob was not modelled as well
  • No ribs, pockets, holes or text — a plain tray

Measured checks

  • watertight (gate)
  • one sub-part named part (gate)
  • seeds the refused reference with segment_reference
  • the seeded mask is accepted and stored
  • applies after seeding are match-scored
  • silhouette matches the seeded reference
  • bbox 174.34×107×17.98
  • the tray is hollow
  • the floor is solid

Ranked by hand

One reviewer put one attempt from each model in order, best first, ties left tied.

  1. #1 tie
    Gemini 3.8 Flash
    Gemini 3.8 Flash's ranked attempt
  2. #1 tie
    GPT-6 Astra
    GPT-6 Astra's ranked attempt
  3. #1 tie
    GPT-5.6 Sol
    GPT-5.6 Sol's ranked attempt
  4. #1 tie
    Fable 5.1
    Fable 5.1's ranked attempt
  5. #1 tie
    Opus 5
    Opus 5's ranked attempt

Ordering by hand decides Human rank and Human rate on the leaderboard; it never affects Overall or Strict.

Every model's attempts

Run baseline-2026-092026-09-10commit 41aec7daspend $339.70 + $48.02 AI judging, 7 unpricedOther runs: 2026-09-04T20-02-10-stage2-hard-cases, 2026-09-04T19-48-31-screen-14-fixed