Depth-Video Character Reskin to Seedance 2.0

A source video's motion is flattened into a temporally-consistent depth video by a video-native depth model, and a single extracted frame is separately re-designed into a brand-new character and setting by an image editor. Seedance 2.0's reference-to-video mode then takes both the depth video and the new character image together and re-renders the original motion performed by the new character in the new world.

Properties

family: style-control
modality: video -> depth video + edited reference image -> re-rendered video
whyMultiModel: Video Depth Anything preserves temporal geometry (every limb movement and camera move) but discards all appearance information; a separate image editor is needed to author an entirely new character/scene as a still reference, which a depth model cannot generate; and Seedance 2.0's own joint video+reference conditioning is required to fuse the two into one physically coherent render, since neither the depth model nor the still-image editor can produce video output on its own.
steps: [object Object], [object Object], [object Object]
controls: model_instructions / video_style text guiding the character-edit step; frame_number selecting which source frame becomes the edit reference; depth video resolution/colormap; number and mix of reference assets fed to Seedance 2.0 (up to 12 images/videos/audio per its docs).
exampleStack: Video Depth Anything (video -> depth video) -> FLUX.2 [klein] 9B edit (new character/scene still) -> Seedance 2.0 reference-to-video (depth video + still -> re-rendered clip)
useCases: [Dance and motion-transfer content that reskins a performer without any pose-estimation rig][Reskinning existing footage with a wholly new, freshly generated character][Style/genre conversion of live-action motion while keeping exact choreography and camera movement]
pitfalls: Depth-only conditioning discards face, outfit, and location by design, so if the edited reference image's body proportions or scale disagree with the depth video, limbs can warp or misalign on re-render; very fast or complex motion can outrun what Seedance 2.0's reference-to-video mode can faithfully reproduce from a depth signal alone.
source: fal.ai Workflow Templates: depth-to-seedance (Styling video using Video Depth Anything and Seedance 2.0)

Provenance

Source: https://fal.ai/workflows/templates/lovis/depth-to-seedance

Depth-Video Character Reskin to Seedance 2.0

Properties

family

style-control

modality

video -> depth video + edited reference image -> re-rendered video

whyMultiModel

Video Depth Anything preserves temporal geometry (every limb movement and camera move) but discards all appearance information; a separate image editor is needed to author an entirely new character/scene as a still reference, which a depth model cannot generate; and Seedance 2.0's own joint video+reference conditioning is required to fuse the two into one physically coherent render, since neither the depth model nor the still-image editor can produce video output on its own.

steps

[object Object], [object Object], [object Object]

controls

model_instructions / video_style text guiding the character-edit step; frame_number selecting which source frame becomes the edit reference; depth video resolution/colormap; number and mix of reference assets fed to Seedance 2.0 (up to 12 images/videos/audio per its docs).

exampleStack

Video Depth Anything (video -> depth video) -> FLUX.2 [klein] 9B edit (new character/scene still) -> Seedance 2.0 reference-to-video (depth video + still -> re-rendered clip)

useCases

[Dance and motion-transfer content that reskins a performer without any pose-estimation rig][Reskinning existing footage with a wholly new, freshly generated character][Style/genre conversion of live-action motion while keeping exact choreography and camera movement]

pitfalls

Depth-only conditioning discards face, outfit, and location by design, so if the edited reference image's body proportions or scale disagree with the depth video, limbs can warp or misalign on re-render; very fast or complex motion can outrun what Seedance 2.0's reference-to-video mode can faithfully reproduce from a depth signal alone.

source

fal.ai Workflow Templates: depth-to-seedance (Styling video using Video Depth Anything and Seedance 2.0)