Depth
Recovers a sense of near and far from one photo, with no second camera and no depth sensor — the same trick you use when you close one eye and can still tell what's in front.
No depth map yet
Add a photograph with clear foreground and background — a corridor, a street, a room. Flat, head-on subjects give the model very little to work with.
These are relative values between 0 and 1. The model can tell you that one region is nearer than another; it cannot tell you that anything is a particular number of metres away. Converting to real distance needs a known reference in the scene — a calibration this model never receives.
The model was trained on millions of images paired with known depth, so it learned the cues humans use unconsciously: things higher in the frame are usually further away, textures get finer with distance, known objects have known sizes. It outputs one relative distance value per pixel.
Monocular depth is fundamentally ambiguous: infinitely many 3D scenes project to the same 2D image. The model resolves that ambiguity with learned priors, which is why it fails confidently on scenes unlike its training data.
- The output is relative, not metric. It tells you A is nearer than B; it does not tell you A is 2.4 metres away.
- Reflections, glass and mirrors break it — the model reads the reflected scene as real geometry behind the surface.
- Flat printed images of scenes often produce a full depth map, because the model recognises the depicted scene, not the paper.
This page says “the model predicts”, not “the model knows”. That is deliberate. None of these systems understand the images or sentences you give them; they map inputs to outputs using patterns fixed at training time. The difference matters most exactly when the output is impressive.