A normal map generator that starts with one photograph

A conventional normal map generator converts a height or bump field you already own into per-pixel surface orientation. That is the right tool when a sculpt, a scan, or a displacement pass produced the height field, and it is the reason the free browser utilities that dominate this category ask for a height map at all. It leaves you stuck one step earlier: you are holding a photograph, no height field exists yet, and there is nothing for those tools to convert.

This page covers the other direction. The picture is the input. DepthCarve estimates relative geometry from that single frame and derives the orientation from the same estimate, so the entry point is the image in hand rather than a field you would first have to invent by hand.

The estimate is built from monocular cues: perspective, occlusion order, relative object size, texture gradient, and the way light falls across a surface. Mirrors, glass, polished metal, and water return a scene the camera is not pointed at. Flat untextured walls offer no gradient to read, fine repeating texture can fold into itself, and thin structures such as wires or hair fall between sampled regions and arrive broken or merged with their background.

Cast shadows are the expensive error, because a hard shadow is dark and sits directly against the object it describes, which is also the pattern that signals proximity. Nothing in a single frame separates the two, so the shadow may arrive as a ridge. Hidden sides and interior cavities are likewise unrecoverable, which is why the output is best treated as front-facing surface information rather than a reconstruction of the whole object.

A direction stored per pixel means nothing until the space and the channel order are known. Provider output may be expressed in camera space rather than the tangent space a material expects, and the green channel differs between OpenGL- and DirectX-style pipelines. Neither convention is wrong; connecting an unchecked map to a shader and adjusting the material until the lighting merely looks acceptable is what consumes the afternoon.

Verify four things against a known surface such as a lit sphere: axis orientation, channel order, handedness, and whether values are object-, tangent-, or camera-relative. A sphere or a cube answers that faster than a full render. The versioned manifest in the production bundle records what was produced, so the check is repeatable instead of being reconstructed from memory.

The approved job returns a bundle rather than one file. Depth and normals come from the same estimate, which is what makes a later seam diagnosable: when the two are produced by unrelated preprocessors, a crease in one can sit a pixel away from the crease in the other and the shading artefact belongs to neither. The bundle also carries raw floating-point depth, the true 16-bit forward and inverse pair, a watertight mesh, and a parameter manifest, with a point cloud, GLB, and camera data when the provider supplies them.

Provider asset links are temporary, typically about an hour, and DepthCarve does not keep the ZIP. Move the bundle into the project archive as soon as it downloads. One approved job costs one credit, a failed provider call returns its debit automatically, and re-encoding the same approved job never charges twice.