Nobigi

Nobigi (Neural optimized baked interactive global illumination) explores how to replace heavy baked lightmaps with tiny neural networks that run directly in shaders. The result is a scene that looks like traditional baked GI but responds smoothly to dynamic inputs such as light rotation. The full demo lives at jure.github.io/nobigi, with background in the blog post.

Dataset and baking pipeline

The training set is generated from Blender using rotate_light_and_bake.py:

• A target scene (the coffee cup) is baked at many light orientations over two axes.
• Each baked frame stores diffuse direct + indirect lighting for a surface, producing a dense grid of lightmap images.
• Filenames encode light rotations, which become the dynamic inputs for training.

Neural representation

The Colab notebook trains a compact SIREN-style MLP per surface:

• Inputs: x, y, and two light-angle parameters (4D total).
• Outputs: r, g, b values for that surface.
• Sine-activated layers with weighted skip connections keep the model small (a few thousand parameters) while preserving high-frequency detail.

GLSL export and runtime

The trained weights are converted into GLSL with model_to_shadertoy, packing parameters into vec4/mat4 blocks to avoid shader compilation bottlenecks. At runtime:

• Each pixel runs inference in the fragment shader.
• Lighting smoothly interpolates between baked states (no stepping or video textures).
• The shader payload stays in the tens of kilobytes, far smaller than video lightmaps.