Vision node
Depth Estimation
On-device monocular depth estimation (Depth Anything V2 or V3): outputs a grayscale depth map from any connected image — near is bright, far is dark.
Inputs
| Port | Type | Default |
|---|---|---|
| ImageWhat to estimate depth from: any image cable — a Camera Input, a Video Player, an Image node, any generator. The model sees the whole frame and guesses how far every pixel is. | image | — |
| EnableOff stops the model entirely (no Neural Engine work) and clears the node's state. On resumes — the first result appears after about a second of model warm-up. | bool | true |
| ModelWhich Depth Anything generation runs: V2 Small is roughly twice as fast (≈30fps) — the live default; V3 Small (≈15fps) reads geometry more faithfully and adds finer detail. Two nodes with different choices on the same image compare them side by side. | int | 0 |
| Depth FPSCaps how often the model runs. 0 = Auto: as fast as it finishes (≈30fps on Apple Silicon). A lower cap saves power — the map simply updates less often; the image output itself stays at full frame rate. | int | 0 |
| XHorizontal centre of the depth picture in Grafit units (0 = the middle of the view). | float | 0 |
| YVertical centre of the depth picture in Grafit units. | float | 0 |
| ScaleSize multiplier of the depth picture, 1 = the source's size. Not clamped; 0 draws nothing. | float | 1.0 |
| RotationRotation of the depth picture in degrees, clockwise positive. | float | 0 |
| MirrorHorizontal flip — match a mirrored (selfie) camera so the depth map lands on the image. | bool | false |
| Flip YVertical flip, for the odd source whose orientation disagrees. | bool | false |
| InvertSwaps near and far: by default near is bright and far is dark; inverted, distance becomes brightness — handy when driving fog or blur amount directly. | bool | false |
| OpacityThe map's opacity on the Canvas. The depth values themselves are unaffected — this only fades the drawn image. | float | 1.0 |
Outputs
| Port | Type |
|---|---|
| ImageWhat to estimate depth from: any image cable — a Camera Input, a Video Player, an Image node, any generator. The model sees the whole frame and guesses how far every pixel is. | image |
| Width | float |
| Height | float |
How It Works
Depth Estimation runs monocular depth on the Neural Engine, entirely on your Mac. Cable any image — a camera, a video, a generator, a whole Canvas — and out comes a live grayscale depth map: near is bright, far is dark, at roughly 30 ms a frame.
The Model dropdown offers two generations of Depth Anything: V2 Small · Fast (the live default, around 30 fps) and V3 Small · Quality (around 15 fps — ByteDance's Depth Anything 3, with finer geometry and more faithful form). Two nodes with different choices on the same cable compare them side by side. Whatever the variant, the output speaks one language — the render path never knows which model ran.
Around the model sit the ordinary generator controls (Y, Scale, Rotation, Mirror, Flip Y, Invert, Opacity) so the map can be framed and oriented to match its source. While the model warms up — the first-ever run compiles for the Neural Engine — a sweeping band says working on it rather than showing nothing.
Both models are Apache-licensed, and their licenses ship in Settings ▸ About.
Use Cases
- Camera-driven focus — Into Depth of Field: a live rack focus that follows whatever is actually nearest.
- Atmosphere — Into Fog: a real camera feed gains distance-graded weather.
- Paint that follows volume — Into Impasto's Depth Flow: strokes turn along iso-depth lines and wrap faces the way a painter does.
- Distance masks — Into Depth Adjust's Band mode: carve out one exact slice of distance as a mask for anything downstream.
- Displacement — A depth map is a heightfield; drive a Geometry Noise or a particle field with it.
Tips
- The map is relative depth, not metres. Use Depth Adjust to map it onto whatever range the consumer wants.
- Cap the rate with the Depth FPS setting when full speed is not needed — the Neural Engine costs power.
- V3 is the better picture; V2 is the one to perform with. Switching models does not change anything downstream.
- Mirror matters when the source is a self-view camera — keep the depth map in the same left/right sense as the picture it describes.
- Invert if your consumer expects far-is-bright; Depth Adjust can do the same and more.
Open Depth Estimation in the interactive catalog →
Depth Estimation is one of 381 nodes in Grafit Pro.
Free download. Every feature for 7 days, no card. Then €99.99 once — or keep playing your projects as a Player.