Photo to 3D: the five approaches, compared honestly

1. Image generation: a new picture, not a model

Diffusion models can produce a convincing image of your room restyled, or your ring from another angle. Nothing underneath is 3D: there are no walls or prongs to measure, and asking for a second change produces a different picture rather than an edit of the first. Good for mood boards; not for anything you will measure, build or manufacture.

2. Depth estimation: how far away each pixel is

A single-image depth model guesses the distance of every pixel from the camera. The result is a 2.5D relief — convincing from the original viewpoint, empty behind every object, and with no absolute scale unless something in the photo provides it. Useful for parallax effects and as one input to reconstruction; not a model you can walk around.

3. Photogrammetry, NeRFs and Gaussian splats: a surface from many photos

With dozens or hundreds of overlapping photos, these methods recover the visible surface of a real object or space accurately — the best choice when you need the exact shape of something that already exists, such as a heritage object or a scanned room. The output is a dense mesh, a radiance field or a cloud of splats: heavy, without parts, and hard to edit. Shiny metal and gemstones, thin prongs and plain white walls are exactly where they struggle.

4. AI-generated meshes: a plausible shape from one image

Newer generative models produce a textured mesh from a single photo. They are fast and impressive on organic shapes, and the hidden side is invented to look right. The mesh is still a single surface with no idea which part is a leg or a band, sizes are arbitrary, and the geometry is often not closed — a problem for printing, casting or calculating volume.

5. Parametric reconstruction: the object as parts with dimensions

This is Kubikly's approach. A vision model reads the photo or plan and describes the object as structured data — walls, doors and rooms; a band, a setting and stones; a top, legs and panels — and deterministic code builds closed, measurable geometry from that description. The result can be edited by changing a number (“make the table 220 cm”, “six prongs”), volumes and weights can be calculated, and every value says whether it was measured, given by you, calculated or estimated.

Side by side

Where parametric reconstruction is the wrong tool

It rebuilds the type of object it recognises, from a library of parts. A freeform sculpture, a car body, a face or an exact scan of a specific worn object are better served by photogrammetry or a scanner. It also cannot see what the photo does not show: a hidden setting or the inside of a wall is inferred from typical construction and labelled as assumed.

Getting the best result from a photo

• Shoot straight on and from a slight angle; for jewellery and furniture, two or three angles help more than one perfect shot.
• Give one real measurement — a wall length, a ring size, a table width. It calibrates everything else.
• For floor plans, upload the drawing with its dimension lines; scale is then read and cross-checked, not guessed.
• Correct anything wrong by saying so; the change is an explicit, undoable edit.