alakazam Environment Transfer · Preview

One demonstration, many worlds.

Egocentric manipulation episodes re-rendered into new environments while task motion, hands, and object identity stay locked. 9 episodes across 6 environments, 25 clips. Structure-locked video restyle with depth and edge control.

25
clips
9
episodes
6
environments
150 @ 30 fps
frames
0.87
structure gate · mean edge-SSIM
Clips

Each clip pairs the source episode on the left with the restyled result on the right. Hover or tap a clip to play.

METHODOLOGY

Source episodes come from Apple's EgoDex dataset, a public research release. Each clip is re-rendered with a structure-locked video restyle under depth and edge control, which holds the recorded hand and object geometry while replacing the surrounding scene.

Every clip runs per-clip gates before it is kept: frame-count parity against the source and a structure-similarity check on edges. Across this set the structure gate holds a mean edge-SSIM of 0.87 (range 0.75 to 0.96). There is no audio. Demo resolution is 960 x 544 per pane, 30 fps.

Pick and place runs 126 frames; every other episode runs 150 frames. Seeds are recorded per clip so any result can be reproduced exactly.

Source episodes: EgoDex (Apple), CC BY-NC-ND 4.0. This page is an internal research evaluation shared for review. Not a redistribution channel; production augmentation runs on partner-owned data.
Attribution: github.com/apple/ml-egodex