Depth videos, not raw clips, make character swaps cleaner in Seedance 2.0

Sniper_yoha · reddit · 2026-07-21

A creator says the key to clean character swaps in Seedance 2.0 is not a better prompt, but converting the reference clip into a **depth video** first. Using the raw clip as reference tends to make the model cling to the original person and setting, which makes swaps muddy. By running the clip through **Depth Anything V2** frame by frame, the author keeps motion structure, timing, weapon path, and camera movement while stripping away most of the original appearance. That makes it much easier to drop in a new character and a new location. The test case is intentionally silly: a short martial-arts reference clip with a Guan Dao form, rewritten into a synthetic character — a 50-year-old market auntie — in a street market setting. The result depends on two practical rules: - Keep the reference under 15 seconds so depth remains stable - Make the prompt strongly enforce prop continuity, so the Guan Dao never flickers away or morphs into another weapon The author also notes that they built a small local depth-conversion tool with a Gradio UI and Depth Anything V2 on the backend, running on the same endpoint as Seedance 2.0. They remind readers that the choreography and music in the source clip still belong to the original rights holders.

Original post →

More from coding & agent

coding & agent channel →