Depth Estimate
Estimate depth from a single image with machine-learning inference.
What it does
Depth Estimate reads a single image and predicts a depth map from it with an ML model, needing no camera, sensor or 3D data. The result is a greyscale image where brightness encodes distance: by default dark is near, light is far.
Quality picks the model size: fast is quickest, accurate takes longer and resolves finer detail. Metric switches the model to output depth in real metres instead of a relative near/far range. Invert flips the brightness mapping so near reads as white.
When to use it
A flat photo or video frame needing a depth map to drive a 2.5D parallax push with Displace
A depth-based mask (near/far selection, fog falloff, focus falloff) without manual painting
A depth signal into a Color Map node for a heat-map style visualisation
Real-world distances rather than relative near/far, with Metric on
| Parameter | Type | Default | What it does |
|---|---|---|---|
image | Raster | — | Image to estimate depth from (any Raster source) |
quality | String | "fast" | Quality: fast (~25ms), balanced (~60ms), accurate (~150ms). One of: fast, balanced, accurate. |
metric | Boolean | false | Use metric depth model (absolute depth in meters) |
invert | Boolean | false | Invert depth (near=white instead of near=black) |
enabled | Boolean | true | Enable depth estimation |
position | Vec3 | (0, 0, 0) | Position offset in world units (X, Y, Z) |
strength | Number | 1 | Effect strength / wet-dry (0 = no effect, 1 = full). Editable, wirable, keyframable. |
Gotchas
Depth is estimated per image, with no notion of a previous frame, so on video the depth map can flicker slightly frame to frame rather than tracking smoothly as a real depth sensor would.
Where a downstream effect expects near-white instead of near-black, use Invert rather than a separate levels flip afterwards.
Accurate quality can take roughly 5–6 times longer per frame than Fast. Use it on stills or final renders, not on a live feed being scrubbed in real time.
Worked example
Add a Depth Estimate node and wire an image into Image.
Leave Quality on fast while setting things up, and switch to accurate for the final look.
Wire depth_map into Displace to push pixels toward camera by estimated depth, creating a 2.5D parallax effect.
If the parallax reads backwards, turn on Invert to flip near/far.
See also
Normal Estimate · Object Detect · Vision Segment · Rasters (images)