TL;DR
Side-by-side MANO playback of GT vs four baselines
(MDM, MoMask, MotionMillion, Kimodo)
on 9 examples (3 per train / val / test split) re-rendered at
5 caption granularities each — 45 distinct text prompts in total.
L0–L4 = caption detail; Prev / Next or arrow keys to step samples.
What you're looking at: each sample shows 5 hand viewers — GT (the captured ground-truth motion) plus four generative baselines (MDM diffusion, MoMask masked-token, MotionMillion scaled AR, Kimodo). The viewer carries 9 examples × 5 caption levels = 45 prompts, rendered as 189 MANO sequences in total (9 GT + 9 × 5 × 4 baseline outputs). Initial decode + WebGL warm-up takes 2–5 s; only the visible card animates.
- Caption level — click an L0..L4 button at the top to swap every method's output to the prompt of that length. GT is identical across L0–L4; only the baselines change.
- Sample navigation — use Prev / Next, the dots, or ← / → arrow keys to step through samples in the active split (train / val / test).
- Per-cell controls — each viewer has its own play / pause / scrub and orbit camera (drag to rotate, wheel to zoom).