Video Depth: Extract Spatial Structure from a Reference Video
Turn a reference clip into grayscale depth footage. Learn how to choose settings, recognize the result node and export it.
A depth video represents predicted relative distance through a sequence of grayscale frames. Use it to inspect silhouettes, occlusion and perspective changes when preparing spatial references for a creative project.
Compare a real result
The players below show the source and its completed depth output. Playing, pausing, seeking or changing playback speed in either player synchronizes the other. Pause and scrub through the clip to examine contours against the original.
Open the original video · Open the depth result
Both files in this example are 600×400, 24 fps, 141 frames and 5.875 seconds. These measurements describe this sample, rather than a guarantee for every input or output.
Extract depth on the canvas
- Upload a video or select an existing video node. Wait for the upload to finish.
- Click Extract Depth in the toolbar above the node.
- Choose an output resolution. Start with Auto; available choices come from the current model.
- Wait for the points estimate, then start processing. No prompt is needed. A single available model is selected automatically; a selector appears when several are available.
- A connected result node appears to the right of the original. When processing finishes, preview or download the video.
This walkthrough uses simulated points and local task states. It never submits a real generation or charges points. Its completed state displays the saved sample above: changing resolution in this demo does not re-encode that sample. Actual cost and timing come from the live canvas quote and task.
Recognize the result
The default title starts with Extract Depth ·, followed by the source title. A dedicated icon appears in its upper-left corner and type displays. The original remains on the canvas. Rename the result if useful, follow its connection to the source, and use the existing video playback, preview and download actions.
Interpret the grayscale footage
Check the brightness direction against known near and far objects in the original before comparing movement. Brightness represents predicted depth, not object color, source luminance or an exact distance in meters. Faces and clothing geometry may remain recognizable; depth output is not an anonymization method.
Settings and recovery
Resolution controls output size, not motion strength. Evaluate a short clip with Auto before requesting other sizes. If a quote fails, read the inline message and retry the quote; processing remains unavailable until it succeeds. A failed task can be retried from its result node. The demo includes a simulated failure so you can explore that state safely.
If the toolbar entry is missing, confirm that the video has finished uploading and a depth model is available to your account. For repeated failures, record the clip, settings and error rather than repeatedly submitting the same task.
Boundaries
The current service processes at most 2,400 frames. Test short clips first and do not assume an over-limit output covers an entire long input. Fast movement, occlusion, transparent or reflective surfaces can produce errors. Flicker and scale drift can occur. Depth footage does not provide skeleton keypoints, action labels, camera trajectories or a 3D reconstruction.
Continue reading
Director Workbench: Camera Blocking, Capture Preview, and Send Back
Use the Director 3D node to block actors, preview cameras, review saved frames, and confirm destructive actions before moving references back to the main canvas.
Video Parameters — Resolution, Duration, Model & More
Configure video generation settings including resolution, duration, audio, model variant, and advanced options.