raymondyang.dev
← map
Recorded output of the real models. Hover a card.
7 s of the BANDAI Card Games Fest 25-26 Las Vegas finals table (official broadcast, cropped to the table), with the detections of Kenbunshoku's models, recorded every 0.2 s. Physical play is the hard case.

Try the real model

The toolkit's demo page runs these exact models in your browser. Pick One Piece, drop in a screenshot or a clip of play, and point at a card. Nothing is uploaded.

Four days, one rented GPU

4 daysfrom starting on One Piece to release on the Chrome Web Store
$13.99of GPU time: 15.7 h on one rented RTX 4090, two training rounds
23labeled real frames, plus 8,000 generated scenes per round, for a catalog of 5,502 cards

Held-out results

ModelsNamed rightWrong of shown
Riftbound weights, One Piece bank31.4% (23.7–42.5)10.6% (4.2–28.0)
Fine-tuned with the adapt guide74.4% (50.3–83.8)3.9% (1.6–11.8)
Fine-tuned + geometric check83.8% (62.5–91.8)2.7% (0.6–10.3)

25 frames from 13 public videos in 9 recording groups, none used in training. 95% intervals over recording groups, so wide. Measured in Python, not in the browser. Most users get the fine-tuned models alone: the geometric check runs only where WebGPU is available, and turns itself off where it is too slow.

Where it struggles

  • Physical tables: 68 of 124 faces named right, against 153 of 173 on the digital client (models only)
  • Official card images carry a "SAMPLE" stamp that real cards don't
  • A whole frame with the geometric check took 1.07 s at the 95th percentile on an M3 Max with WebGPU