Depth Anything V2 motion maps survive character swaps in a Seedance 2.0 test

Few-Profession421 · reddit · 2026-07-28

A user tested a pipeline that turns a short climbing/parkour reference clip into a Depth Anything V2 depth map, then feeds that motion reference plus a newly generated synthetic character into Seedance 2.0.

The result: wall contact, weight shifts, and full-body motion reportedly carried over cleanly, which is usually where character-swap workflows break down. They also say the local depth converter itself was built with GPT 5.6 Sol, and expect the same setup to transfer to dance and martial arts clips.

Original post →

More from Multimodal

Multimodal channel →