What is a normal map?

How a normal map works

A normal is the direction a surface faces, pointing straight out of it. Lighting depends on it: a surface that faces the light is bright, and one that turns away is darker. On a plain low-poly model, the normal changes smoothly across each triangle, so a flat panel lights up flat.

A normal map replaces that smooth normal with a new one for every pixel. The geometry stays exactly the same, but the shading now tilts pixel by pixel, so rivets, pores, wood grain or brick joints catch the light like real relief.

How a direction becomes a colour

Each part of a direction runs from -1 to +1, but a colour channel only runs from 0 to 1. So the value is squeezed: colour = direction × 0.5 + 0.5, and the shader reverses it when it reads the map. The glTF 2.0 spec spells it out: red maps to X, green to Y, blue to Z.

Why normal maps look blue or purple

On most surfaces, most pixels point almost straight out. "Straight out" is X = 0, Y = 0, Z = 1, which encodes to the colour (0.5, 0.5, 1.0), or about (128, 128, 255) in an 8-bit image: a light blue-violet. Only slopes shift the colour, so edges facing one way turn pinkish (more red) and edges facing up turn greenish (more green).

That is why a normal map does not look like a photo of the surface. It holds data, not colour, so it must be loaded as Non-Color (Blender) or linear data, never as sRGB. Loading it as a colour image is the most common reason a normal map shades strangely.

Tangent space vs object space

The direction in each pixel has to be measured relative to something. There are two common choices:

Tangent-space maps are the default almost everywhere. Blender’s manual calls them the most common type because they survive object transforms and mesh deformation. glTF supports only tangent-space normal maps: if a GLB has no tangents, the viewer computes them with the MikkTSpace method from the mesh and its UVs.

Because tangent space is built from the UVs, a tangent-space normal map only fits the UV layout it was made for. Re-unwrap the model and the old map no longer lines up. If UVs are new to you, see UV unwrapping explained.

OpenGL vs DirectX normal maps

There are two conventions for the green channel, and they are exact opposites:

• OpenGL (Y+, "green up"): Blender, Unity, Godot, three.js and the glTF format. The Blender glTF manual lists glTF normal maps as "tangent space, +Y up".
• DirectX (Y-, "green down"): Unreal Engine and some other DirectX-based tools.

Red and blue are the same in both. If bumps look pushed in when they should stick out, or the light seems to come from below, the map is in the other convention. The fix is to invert the green channel, which most texture tools and our normal map generator can do with one switch.

Normal map vs bump map vs displacement map

A bump map only says how high each pixel is, so the renderer has to work out the slope from neighbouring pixels. A normal map stores the slope directly, in two directions, which gives more control over how light bends. Unity’s manual describes both as kinds of bump mapping.

The big limit of both: the edge of the model does not change. Look at a normal-mapped brick wall from the side and its outline is still perfectly flat. When the silhouette has to change, or the detail must be printed, you need real geometry or displacement.

How normal maps are made

• Baked from a high-poly model. Artists sculpt a detailed version, then bake its surface directions onto the low-poly game model’s UVs. This gives the most accurate result.
• Made from an image. A tool reads the texture’s brightness as height and turns the slopes into a normal map. It is fast and works well on stone, wood, fabric and other textures where light and dark follow the relief, but it is a guess from shading.
• Painted or sculpted directly in texturing apps that paint height or normal detail as layers.

Make a normal map in your browser

From a texture or photo: our free normal map generator turns an image into a tangent-space normal map, with strength and detail settings, an OpenGL / DirectX switch and a moving-light preview. The image is processed on your device.

For a whole 3D model: in the AI 3D Texturing Studio, first paint the model’s colour with AI, then press Make PBR maps. The studio builds a normal map from the painted texture (with a Bumpiness slider), plus roughness, metal and ambient-occlusion maps. While the PBR preview is on, the GLB you download carries the normal map in glTF’s normal slot, and PBR maps (PNG) saves it as a separate image. Like any image-based normal map, it is estimated from the texture’s light and dark areas, not baked from extra geometry.

To see how the normal map fits next to the other maps, read PBR maps explained.

Questions