[Lightmap] Bake Tool

[T-rvw]

• Overview
  • The first step is to bake Direct/Indirect Lighting Diffuse. Direct Lighting Diffuse should be same with engine realtime rendering by using same render pipeline. Indirect Lighting Diffuse is the key part to improve lighting visual result.
  • Consider material's AlbedoMap + NormalMap.
  • CPU baker at first. Then consider GPU baker.
  • See Dghelneshi's answer in high_level_overview_of_a_lightmapping_generation which has more professional details.
• Lightmap
  • Simple Lambertian diffuse BRDF without view dependencies
  • Consider NormalMap to avoid flat ambient issue
  • Learn about different BakingMode
• Reference
  • MJP Blog
  • Radiosity Normal Mapping
  • MJP's sig 2015 course
  • MJP's BakeTool
  • baking-artifact-free-lightmaps
  • sig 2013 Lightmap in "The Last Of Us" game
  • CEDEC_Geomerics_ReconstructingDiffuseLighting

[T-rvw]

Lightmap meets with NormalMap to get better surface details in indirect light. Use half life 2's 3 basis's solution to blend irradiance values at runtime.

[T-rvw]

Indirect + Direct Lighting Diffuse:

Direct Lighting Diffuse:

Indirect Lighting Diffuse:

[T-rvw]

@Hinageshi01 You can play with MJP's BakingLab. I got some basic knowledges about baking lightmap implementation.

[T-rvw]

• Generate LightmapUV for multiple scene objects to share one lightmap texture
  • XAtlas Implementation
    • For model's every mesh, do below task in parallel
      • Then it will build mesh faces to different groups in MeshFaceGroups::compute()
        • Loop every face
          • Loop face's all adjacent faces if valid. Create a queue container. Init it with current face.
          • Loop every non-unboundary edge belong to the container's first face(pop up)
            • Query current edge's opposite face. If opposite face's material is not same with face, don't put in same group.
            • Skip if opposite face already put in a group.
            • Record face loop order.
            • Add valid opposite faces to previous created queue container.
      • After split faces into different groups considered connections + same materials, we define group number as UV Charts group count
      • Record invalid faces' information such as indexes.
      • Sort all UV Charts group by every Charts group's face number. RadixSort is used here which is widely-used parallel sort algorithm.
      • For every Charts group, do below task in parallel.
        • ChartGroup::computeCharts
          • For every Chart, parameterize it in PiecewiseParam::computeChart.

[Hinageshi01]

1. Generate UV for all meshes
2. Baking
   • Generate light map
     • UV space rasterization
       • OUT.Position = float4(IN.LightmapUV * 2 - 1, 0, 1);
     • Padding
       • for every empty background pixel, it simply looks at 8 neighbours around it and copies the first non-empty value value it finds.
     • Conservative rasterization, post-padding is not enough for small triangle
       • Repurposing MSAA samples
       • Rendering geometry multiple times with sub-pixel offset √
       • Rendering lines over triangles
     • Optimize shadow leaks
       • push sample points out of shadowed areas where leaks can occur
     • Optimize shadow terminator
       • Adding constant bias to ray start
       • Making shadow rays ignore adjacent faces that are almost coplanar
       • Tessellating/smoothing geometry before ray-tracing
       • Blurring the result
       • Tamy Boubekeur and Marc Alexa. 2008. Phong Tessellation. √
         • Determines if the smooth position intersects another surface
         • If any fragment on a triangle intersects another surface, all fragments on the entire triangle should use the flat position
     • Beware of weird triangles with 2 normals pointing out, and one in
     • Fixing UV seams
       • Not quite sure how this works
   • Trace rays
     • OptiX, RadeonRays, DXR
     • UV GBuffer
       • rasterize surface attributes like above
     • Ray bias, to prevent noisy self-overlapping due to floating-point inaccuracy
       • position += sign(normal) * abs(position * 0.0000002)
3. Some little optimization
   • Denoising
     • Optix AI denoiser
       • Must before UV seam fixing
       • Maybe something like reversible tonemapping is needed to ensure that the input to the denoiser is non-HDR
   • Bicubic interpolation
     • Not quite sure how this works
   • lightmap mip map?

[Hinageshi01]

In game renderer and Radiosity:
http://the-witness.net/news/2010/09/hemicube-rendering-and-integration/
https://web.archive.org/web/20160311085440/http://freespace.virgin.net/hugo.elias/radiosity/radiosity.htm

[Hinageshi01]

Solution from Precomputed Global Illumination in Frostbite
Store L1 SH for every texel.