PurelandFold

PurelandFold is the ground-truth benchmark introduced in FoldingAgent: Inferring Parametric Origami Procedures from Demonstration Videos (Maya Moriya, Sigal Raab, Yael Vinker, Tali Dekel — Weizmann Institute of Science / MIT). It pairs real photos from instructional origami videos with the exact geometric state of the paper at each photographed step.

• Project page: https://maya-moriya.github.io/origami-page

Why this dataset exists

FoldingAgent's task is to infer a parametric, executable folding procedure directly from a sequence of video keyframes — a step-by-step reconstruction of how a flat square of paper turns into a finished origami model. Existing origami datasets provide either static crease patterns or final folded forms, but none pairs real demonstration photos with ground-truth procedural folding states. PurelandFold was curated to close that gap and to serve as an evaluation benchmark for this task.

The scope is Pureland origami — a well-studied subset (Smith, 1980) in which every fold is a simple mountain or valley fold along a straight crease, combined with unfold, rotate, and flip. Pureland keeps the representation compact (every intermediate state is a flat-folded layout of polygonal faces) while still covering a wide range of real instructional folding sequences.

How it was collected

• Sourced from the Easy Origami category on the OrigamiWay website, which offers a clean, diverse set of models; the Pureland-compatible subset was selected.
• All 27 demonstration videos were filmed by the authors themselves (front-facing LED ring light, ambient daylight, iPhone 12, black Bristol board background) rather than scraped from noisy online sources — this avoids filtering/curation bias.
• Keyframes marking meaningful hand-paper interactions were manually extracted from each video.
• Each keyframe was annotated with the full ground-truth geometric state of the paper at that moment: vertex/face geometry, paper topology, crease (mountain/valley/flat/boundary) assignments, per-face front/back orientation, and the layer stacking order.
• Sequences average 12 frames, ranging from 5 to 21.

See Sec. 4.1 and Sec. 5.2.1 of the paper for the full formal state representation and dataset construction details.

Dataset structure

One row per (sequence, step) — i.e., per annotated keyframe. 337 rows across 27 sequences.

column	type	description
sequence	string	Short, normalized name of the folding sequence (e.g. ladybug, gift_card_holder, tulip_stem)
step	int	Keyframe index within the sequence (1 = flat starting square)
image	image	The real photo for this keyframe
cp.svg	string (SVG)	Rendered crease pattern (unfolded 2D layout) for this state
cp.fold	string (JSON)	Crease pattern in FOLD format — vertices, faces, edges, mountain/valley/flat assignment, fold angles, and face stacking order
flat_folder	string (JSON)	Diagnostic output from compiling cp.fold with the Flat-Folder crease-pattern solver (Jason Ku, 2022): folded vertex/face positions, face order, and counts of the taco-taco / taco-tortilla / tortilla-tortilla / transitivity conditions used to certify flat-foldability

All rows in this release come from ground truth (ordinary photographs paired with hand-annotated geometry) rather than from any model's predictions.

flat_folder and Compilation Validity

The paper's primary validity metric, Compilation Validity (CV), checks whether a predicted geometry can be compiled by Flat-Folder into a definite, self-intersection-free folded state — i.e., whether it is syntactically valid, geometrically foldable, and produces a consistent layer ordering. The flat_folder column contains that same compiler's output for every ground-truth state in this dataset (all of which compile successfully, as expected for verified ground truth), so it can be used directly to reproduce or extend that metric.

Citation

If you use PurelandFold, please cite the FoldingAgent paper:

@article{moriya2026foldingagent,
  title   = {FoldingAgent: Inferring Parametric Origami Procedures from Demonstration Videos},
  author  = {Moriya, Maya and Raab, Sigal and Vinker, Yael and Dekel, Tali},
  year    = {2026},
}

License

Released under CC-BY-4.0. Please provide attribution to the FoldingAgent paper and project page above.