Purpose

Video Depth Anything estimates depth from video using temporal history. It aims to reduce the frame-to-frame changes you can get from estimating every image independently.

Use

Place Video Depth Anything under Analyze in the Image view. Select the Video Depth Anything Small model and connect the moving image to in. Use depth to drive focus, displacement or a depth mask.

Turn Normalize off when preserving the model’s temporal depth scale matters. If a downstream effect needs a fixed range, remap the values there.

Inputs and outputs

• in is the source image sequence.
• depth is a single-channel relative depth image.

Controls

• Model selects the Video Depth Anything checkpoint.
• Normalize rescales each frame to 0 through 1 using that frame’s minimum and maximum. It is on by default. Off preserves the raw relative depth values.

Notes

Per-frame normalization can reintroduce scale changes even when the underlying estimate is temporally consistent. Neither setting turns relative depth into metres.

For depth from independent still images, see Depth Anything.

Ports

Inputs

• in image required
  RGBA

Outputs

• depth image
  single channel

Parameters

Model (Modelpath) Str op('video_depth').par.Modelpath

Default:

video_depth_anything_vits.pth

Normalize (Normalize) Toggle op('video_depth').par.Normalize

Rescale each frame to 0-1 from that frame's own min and max. Off publishes the model's raw relative depth. This model is the temporally stable one, and the per-frame rescale is the one step that reintroduces frame-to-frame scale drift, so off is the setting that preserves what this model is for.

Default:

On

Options:

Off, On