Depth Map Generator

What the depth map actually contains

MiDaS v2.1 small estimates relative inverse depth, not distance. A brighter pixel is nearer the camera than a darker one in the same image, but no grey level corresponds to a fixed number of metres, and nothing in the file records scale.

Each map is normalised independently across its own minimum and maximum, so two maps from two photos are not comparable. Mid-grey in a close-up portrait and mid-grey in a landscape describe completely different distances. Generate depth per image and use it within that image; do not expect depth to stay consistent across a sequence of shots.

Inference runs at 256x256 and the result is resampled back to your image's dimensions, so fine structure — hair, wire, foliage, chain-link — comes back softer than the source. The download is an 8-bit greyscale map.

Where it works, and where it breaks

It works best when the photo has genuine depth to recover: a subject standing clear of its background, a product on a surface, a street or interior receding away from the lens.

It breaks down on flat originals — scans, screenshots, artwork, or a subject shot square-on against a wall — because there is no depth cue to find, and the map comes back as low-contrast mush. Glass, mirrors and water are read as the surface itself rather than as whatever sits behind or beneath them.

Using the map for parallax and 3D

The usual job is a displacement or depth channel: drive a 2.5D parallax push in After Effects, displace a subdivided plane in Blender, or mask a depth-of-field or fog pass so it falls off with distance.

It also serves as a control image for depth-conditioned image generators, which expect exactly this kind of relative greyscale map.

Because the values are relative, animate within a single still. A camera move built from one map is stable; cutting between maps generated from different frames is not.