I fold paper for fun. Fold-lab started as a small playground: describe a crease pattern, export a fold trajectory, optionally render it. Along the way I kept bumping into a research-shaped question that is easy to state and hard to measure.

Here is a Blender playback of one exported fold sequence — not a learned model, just the trajectory pipeline moving a sheet:

The awkward bit

From a camera, two sheets can look almost identical while the hidden state is different — valley vs mountain, left vs right diagonal, layer A on top vs layer B, a crease that is “active” vs one that is only drawn. A world model that predicts the next RGB frame has little reason to care. Physics does.

So the failure mode is not “the model is a bit blurry.” It is “the model is blind to a difference that will dominate the next few seconds of motion.”

What I built so far

Mostly scaffolding, not a paper:

• A JS path that turns crease patterns into forward fold trajectories
• Pair generators for deliberately confusing cases (valley/mountain, mirrored diagonals, layer order, micro-offsets, …)
• A rough Hidden Consequence Score (HCS): compare how much a predictor thinks two near-identical starts will diverge, versus how much the true futures diverge
• Blender rendering for demos like the clip above

HCS ≈ 0 means “blind to the hidden difference.” HCS ≈ 1 means “sees it about as strongly as the ground-truth futures differ.”

The early HCS runs are still noisy; some pair types come out invalid under the current matching heuristic. I am not dressing that up as a result.

Why I am writing this anyway

Because the question is clearer than the method. If a model cannot tell valley from mountain when the silhouette matches, then stacking more frames of the same appearance will not fix foldable-sheet prediction. You need either better state, or a predictor that is allowed to switch dynamics when a crease activates.

Fold-lab stays a hobby project. If I later get a stable HCS table on rendered pairs, I will update this note. For now it is a bookmark: origami is a nice stress test for “same pixels, different future.”

---

← Back to blog