Devlog

Every pose gets a figure, drawn straight from its keypoints

All 104 poses render as articulated figures computed from their keypoint data, and the pose pipeline is open source with a methodology paper and a DOI.

  • Poses
  • Research
  • Data

On July 19 every pose in the catalog gained a drawn articulated figure, a shaded mannequin that sits between a stick figure and a human body. The figures are computed directly from each pose's stored keypoints, so all 104 are covered by construction and each one matches its data.

On July 20 the figures gained automatic framing, a fuller torso built along the spine, anatomical proportions, and four color palettes so three- and four-person scenes read as separate bodies. The catalog holds 11 solo poses, 74 for two people, 16 for three and 3 for four.

The same day the figures arrived, the construction pipeline went public as Agaveis/pose-catalog-pipeline under Apache 2.0, with a methodology paper and the concept DOI 10.5281/zenodo.21439174. Paper, code, dataset and site now link in one chain, and the scripts rebuild the published catalog from a fresh checkout.

Under the hood

  • Each figure is inline SVG, rendered with the page and computed deterministically: shaded capsule limbs, ball joints and an ovoid head over the 17 keypoints, with no image model in the loop.
  • Framing fits a bounding box over every visible joint of every figure, grows it by a flesh margin, then scales the scene uniformly and centers it.
  • The torso is a chest, waist and pelvis capsule stack along the spine axis with a minimum width, and limb proportions derive from torso length. Hands are mitts along the forearm and feet are ovoids set perpendicular to the shin, since the keypoint layout has no toe point.
  • Depth comes from the per-joint visibility code: it sets draw order, and occluded geometry is tinted toward the background as atmospheric perspective. Every figure after the first sits on a background halo, and four palettes keep like-roled partners distinct.
  • The public repository holds nine canonical vocabulary files and five scripts that fetch Wikidata positions, build keypoints, assemble dataset rows, build the dataset and render a contact sheet, all self-contained in the repository. The concept DOI covers every version, v1.0.0 has its own DOI, 10.5281/zenodo.21439175, and the archive carries related-identifier links to the dataset and the catalog.
figures drawn from data
104
two-person poses
74
figure palettes
4

See it on the site

Sources and open data

Read the full log