Render to Unwrapped Texture (for baking)

[bhouston]

It is possible to modify the main ThreeJS shaders so that we can enable rendering a final color to an unwrapped texture target. This involves using the world position for lighting, but the desired UV coordinate when writing the final results.

This should be implemented in a fairly flexible fashion to Three.JS.

(There may be an additional pass required after the main render pass to do extrapolation around each rendered component.)

[pailhead]

What manner of abstraction are you looking for. Why is this something not to be done with ShaderMaterial?

[Usnul]

I'm not sure how this would happen. You could do the following for each face:

1. compute a camera matrix for a face normal
2. open up the frustum to 180deg
3. compute transform world position to UV position of a face being rendered
4. set stencil mask to limits of the face UV region
5. render

you would probably end up with some artifacts along the face edges in the map, this could be handled with some filtering.

Another, probably more accurate way, would be to go through the texture, and for each pixel (texel) that is within used UV space (pixel used for some face in the geometry), you would:

1. texel position -> uv (normalize)
2. find face(s) using the uv coordinate
3. compute world position and normal for the uv coordinate
4. set camera transform. Unfold frustum to 180deg
5. render to pixel

you could do some jitter and multi-sampling for each pixel to get good results for edge regions

one extra point that should probably be considered also is margins, you need to render some margins to facilitate mipmaps.

you could combine two approaches, doing first thing for most of the face regions and handle texels that are on/close-to edges with second approach.

[bhouston]

Interest. Have you heard of http://www.cemyuksel.com/research/meshcolors/?

[Usnul]

@bhouston looks interesting, but I don't think i fully undestand it.

Basically you use more UV points than just vertices and map each UV point to a texel explicitly? Then to avoid mipmap artifacts you use different coordinates for each mip level and have custom maps? If i'm not mistaken, i think the paper you are talking about is quite different from the topic of baking, when baking you typically already have a defined geometry and shaders, so it's only about content of the textures, not about addressing scheme. I feel like i'm missing something :D