How projection texturing works: from camera views to a UV texture

The slide projector idea

Picture a white plaster statue in a dark room. Put a projector in front of it and show a photo of a painted statue taken from exactly the same spot. From where the projector stands, the plaster statue now looks painted. Walk around to the side and the illusion breaks: the image is stretched along the sides and the back is still white.

Projection texturing does the same thing in software, with one important extra step: instead of only shining the image, it writes the colour into the model’s texture image so it stays there when you walk around. That writing step is called baking (more in What is texture baking?).

The pipeline, step by step

1. Place a camera and render the model

The tool places a virtual camera around the model, for example straight at the front. It renders what that camera sees. For AI texturing the render is usually plain (grey, or with simple helper images such as edges or depth) because the AI only needs to know the shape and where things are.

2. Paint the view

The render goes to an image AI together with your prompt. The AI returns a painted image of the same view: same outline, same pose, but now with fur, armour or stone. The better the AI keeps the outline in place, the cleaner the projection, because a painted ear that drifts a few pixels will land on the cheek.

3. Find which surface point each pixel belongs to

For every point on the mesh, the software asks: "where does this point appear in the camera’s image?" It runs the point through the same camera maths that drew the render, and gets a pixel position. That pixel’s colour belongs to that point.

4. Check visibility

Not every point the maths lands on is really visible. The back of the head projects onto the same pixels as the face. So the tool keeps a depth record of the render and only accepts points that are the closest surface at that pixel. Everything behind is left alone for another view to paint.

5. Write into the UV texture (baking)

Each visible point has a UV coordinate, an address in the 2D texture image. The colour is written at that address. After one view, the front of the model is painted in the texture and the rest is still empty. If UVs are new to you, read UV unwrapping explained first; projection only works if the model has a clean UV map.

6. Repeat and blend

Move the camera (right, back, left, top) and repeat. Where two views overlap, the tool has to choose or blend. A common rule is to trust the view that looks at a surface most directly, because a surface seen at a sharp angle gets stretched, blurry pixels.

Why seams appear, and how tools reduce them

The most common complaint about AI textures is a visible line where one view hands over to the next. Three things cause it:

• The views disagree. The AI painted the front and the side separately, so the shade of blue or the pattern of scales is a little different. Research such as synchronised multi-view diffusion tackles this by letting the views share information while they are generated.
• Stretching at grazing angles. Near the edge of a view the surface turns away from the camera, so one image pixel covers a large patch of surface. Weighting by viewing angle and feathering the edges of each view softens this.
• Small misalignment. If the painted image shifted slightly compared with the render, details land in the wrong place and break at the border.

In practice, tools combine several fixes: angle-based blending, soft edges, giving the AI the already-painted model as a guide for the next view so it continues the same colours, and seam filling in UV space so the texture does not show thin gaps along UV island borders.

Our studio weights each view by how squarely it faces the surface and matches the colour of each new Auto view to the paint already there, which hides most seams. For a seam that is left, the free Heal brush is fastest: stroke along the line to blend it into the colour around it. For a bigger mismatch, turn the model so the seam is in the middle of the screen and paint that one view again; the seam is then in the centre of a view, where projection is sharpest. The tutorial fix seams with the free heal brush shows both.

Occlusion: the spots no camera sees

Four cameras around a model miss some places: under the arms, between the legs, under the chin, inside an open mouth, the underside of a car. These remain unpainted after automatic texturing. Different tools handle them differently:

• Add more views, from above, below or in between.
• Fill from neighbours, spreading nearby colour into small gaps.
• Let the user aim a camera at the gap and paint only the missing area.
• Repair in UV space with a second AI model that completes the texture image directly (an active research area).

In the AI 3D Texturing Studio, Auto texture can paint 4, 6 or 8 views. With 6 or 8, the extra cameras are picked by coverage: the studio tries many camera positions and adds the ones that see the most of what the first four missed, such as the top and the underside. After that, unpainted areas show as grey. Turn the model so you can see a grey spot, tick "Only fill grey (unpainted) areas", and press Paint this view. Only the empty texels are painted, and the paint you already like stays as it is.

What if the model has no UVs?

Projection needs somewhere to write, which means a UV map. Many files have none: most STL files, many scans, and some exports. A tool can either refuse those files or create UVs for them first. Our studio creates a UV atlas automatically when a mesh has no UVs. When a file does have UVs, Keep my UVs is the default, so the texture fits your own layout and drops into your pipeline. If those UVs overlap or are mirrored, two parts share one area of the texture and get each other’s paint; the studio warns you when more than 5% of the painted area overlaps and offers Make new UVs for a fresh layout where every part gets its own texels.

When the studio makes new UVs, download the GLB rather than the PNG on its own: the GLB carries the new UVs, while your original file does not know about them. The tutorial keep your own UV map when texturing with AI covers the choice.

Projection for one part only

Projection also makes local edits easy. If you brush a mask over part of the model, only the texels under that mask accept new paint. Paint outside the mark stays as it was. That is how the Select region brush works in the studio: change a shirt colour or a face without touching the rest. The tutorial repaint part of a 3D model walks through it.

Projection vs other ways to texture with AI

Many production tools mix these. For a beginner, the key idea is the same in all of them: every colour in the final texture came from a pixel in some image, and projection is the bridge between that image and the surface.

Questions