PeterAM4/chessboard-renders

Chessboard renders

500,000 rendered chess boards, each labelled with its position (FEN) and its four corners. Train a model to read a board from an image, then test it on boards that differ from training in one controlled way: low cameras, unseen piece shapes, unseen colours, metal and glass, harsh light or a cluttered table.

from datasets import load_dataset

ds = load_dataset("PeterAM4/chessboard-renders", streaming=True)  # every split, 129 GB, streamed

# or download a part: 20,000 train images (5 GB) and one test split
part = load_dataset("PeterAM4/chessboard-renders", data_files={
    "train": "data/train/0000[0-9].parquet",
    "test_low_camera": "data/test_low_camera/*.parquet",
})

Splits

split	images	what changes
train	474,000	all factors in their training range
validation	10,000	all factors in their training range
test_iid	2,000	like train, other games
test_low_camera	2,000	low cameras, 25 or more pieces
test_unseen_pieces	2,000	piece shapes never seen in training
test_unseen_colours	2,000	piece colours never seen in training
test_unseen_materials	2,000	metal and glass pieces
test_unseen_lighting	2,000	hard sun, dim or coloured light
test_unseen_background	2,000	patterned tables, objects around the board
test_all_hard	2,000	every hard range at once

Each test_unseen_* split moves one factor outside its training range and keeps the others as in training. test_low_camera uses the lowest training cameras with crowded boards, and test_all_hard moves every factor at once. With 2,000 images, a score near 95% is known to about one point.

factor	training range	hard range
camera	6 to 78 degrees above the board, any side, 20 to 50 mm lens	6 to 15 degrees (also seen in training)
pieces	6 mesh sets and 4 procedural families, half with other proportions	4 mesh sets never used in training
colours	white and black, or pairs of hues away from green	both sides green (hues 79 to 162 degrees)
materials	glossy or matte paint	brushed metal or glass
lighting	soft area lights, warm to cool white	a hard sun, dim light or coloured light
background	a plain table	a wood, stone or tile table with 4 to 10 objects around the board

Columns

• image: the 1920 by 1080 JPEG.
• fen: the piece placement, the first field of a FEN (uppercase is White, the light side).
• corners: the corners a1, h1, h8, a8 in pixels (x right, y down). a1 is a dark square, as on a real board.
• pieces (count) and source: game, or scattered (about 7.5%), the same pieces on random squares so a model cannot lean on how games usually look.
• elevation and azimuth of the camera in degrees (azimuth 0 is White's side), piece_set, material, lighting and background: the factors to group results by.
• render: every other value used to draw the image. The full FEN and the Lichess game, the lens and depth of field, the piece, board, table and light colours, the roughness.

How it was made

One position per game of the Lichess database (January 2015), at a random move, so openings, middlegames and endgames all appear. Splits are separated by game and no position appears twice in the whole dataset. Rendered with Blender 4.2 EEVEE (16 samples, ray traced reflections and refraction).

Limitations

• Every image is rendered: test on real photos too, for example ChessReD.
• The boards have no border or coordinates, and nothing like a hand covers the board.
• The clutter is simple shapes, and every image holds one board.

Licence and credits

The images and labels are CC BY 4.0. The positions come from the Lichess database (CC0). The piece meshes, all free:

• Chess Set by Riley Queen, Poly Haven (CC0)
• Staunton and Readymake Duchamp sets, clarkerubber/Staunton-Pieces on GitHub (MIT)
• Chess by pjhanzlik and Chess Pieces by Clint Bellanger, OpenGameArt (CC0)
• Staunton Full Size Chess Set by Jeyhun1985, Sketchfab (CC BY 4.0)
• Held out for test_unseen_pieces: Minimalist Chess Set by evanlissoos and Low Poly Magnetic Chess Set by grantisimo, Thingiverse (CC BY), chess3D by JoeMacibi on GitHub (MIT), 3D Chess Pieces by tunakron, OpenGameArt (CC0)

Citation

@misc{massih2026chessboard,
  author       = {Peter A. Massih},
  title        = {Chessboard renders: 500,000 synthetic chess boards with
                  out-of-distribution test splits},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/PeterAM4/chessboard-renders}}
}