Seedance 2.5 uses depth maps to preserve actors’ motion in scene edits

• Filmed performances supply the action: CuriousRefuge’s cliff test transforms a simple greybox shoot at Promise into a cinematic cliffside scene.
• Look and motion get separate references: the cliff prompt assigns an edited still to the first frame and a depth video to motion in Seedance 2.5.
• Faces and costumes get additional inputs: the cave workflow includes the original footage for facial details and a separate outfit image.

The full tutorial also transforms a two-person studio conversation into a fancy bar scene. The depth processor on fal costs $0.04 per second of footage, making a ten-second depth pass $0.40 before image or video generation.

Curious Refuge filmed an actress running in a plain studio environment at Promise, then generated the cliffside setting around her performance.

For the still-image edit, Curious Refuge names GPT Image 2.0 in its cliff walkthrough. Two images establish the shot:

• @img1: the starting frame from the greybox footage.
• @img2: a reference for color grading, lighting and style.

The image-editing prompt was:

The clever input is a black-and-white depth video of the original take. Curious Refuge generated it with fal’s Depth Anything Video, the tool linked in both threads.

The service runs Video Depth Anything, a CVPR 2025 model, and produces per-frame depth estimates with temporal consistency across frames.

The cliff generation uses three assets, according to the tutorial:

1. Edited starting image: establishes the first frame and visual treatment.
2. Depth video: supplies the motion reference.
3. Original live-action footage: carries the performer’s likeness.

Curious Refuge assigns the depth clip to @Video 2 in the published cliff prompt:

Generated cliff scene

The demonstrated Seedance interface runs through Magnific, with these settings described in the tutorial:

• Resolution: up to 1080p.
• Duration: up to 30 seconds, with the output duration set to match the source clip.

A second studio take starts with an actress picking up a screwdriver. The generated scene replaces it with an ancient scroll and moves the action into a cave.

For this starting-frame edit, Curious Refuge names GPT Image 2.5. Its prompt specifies the prop replacement, dirt ground and torchlight:

Curious Refuge supplies four assets in the cave generation step:

1. Edited starting frame: establishes the cave and scroll.
2. Depth video: comes from a depth pass on the original cave take.
3. Original greybox footage: fills in facial details.
4. Outfit image: supplies the character’s new clothing reference.

The cave prompt uses a different video slot from the cliff example:

Curious Refuge also documents departures from the source performances and references in its full tutorial:

• Depth is optional: the presenter says a depth map can sometimes make the result worse.
• Foot contact: the cliff version using only the original video and edited image has slightly unrealistic contact between the runner’s feet and the ground.
• Dialogue performance: the bar conversation contains subtle differences from the actors’ original performances.
• Hand movement: another bar generation extends the actress’s reach farther than the source footage.
• Costume fidelity: the cave result omits the hat from the outfit reference.
• Lighting: a flat-lit source take produces a bar result the presenter also finds too flat; the tutorial’s image-model comparisons show different amounts of relighting.