Lumbre

A biased GPU path tracer for Apple Silicon, written in Odin + Metal.

Inspired by Redshift — trades physical accuracy for speed using irradiance caching, photon mapping, and aggressive bias controls. Targets 10-100× faster convergence than unbiased renderers.

Status

Stage 4A + 4B complete — Pure-Odin EXR writer (FP16, multi-layer) and --aovs for multi-AOV EXR output. Builds on the Stage 3 Principled BSDF and Stage 2 biased GI.

Quick start

odin build .
./lumbre --scene assets/cornell_box.obj --width 1920 --height 1080 --spp 100 --output frame.exr

Or with AOVs:

./lumbre --scene assets/cornell_box.obj --width 1920 --height 1080 --spp 100 --aovs --output frame.exr

Or use the test script:

scripts/render_cornell.sh --quick

Usage

lumbre [options]

  --scene, -s <file.obj>     Load OBJ scene
  --width, -w <int>          Image width (default 1024)
  --height, -h <int>         Image height (default 576)
  --spp <int>                Samples per pixel (default 50)
  --depth <int>              Max bounces (default 20)
  --max-radiance <float>     Firefly clamp (default 1000)
  --roughness-cutoff <float> Bias: treat rough Principled as diffuse (default 0.95)
  --glossy-bias <float>      Bias: damp Principled roughness toward mirror (default 0)
  --output, -o <file.png|exr> Output file (default render.png)
  --aovs                     Write AOV layers to .exr output
  --cpu                      Force CPU renderer
  --gpu                      Force GPU renderer
  --quality <preset>         Quality preset: draft, preview, final
  --gi-dist <float>          Cache lookup distance (default auto)
  --photon-radius <float>    Photon search radius (default auto)
  --debug <mode>             Debug: 1=albedo, 2=normal, 3=depth,
                             4=primitive id, 5=direct, 6=light count,
                             7=direct candidates, 8=shadow visibility,
                             9=indirect, 10=GI cache hits,
                             11=photon contribution,
                             12=GI cache samples,
                             13=GI cache confidence

Without --scene, renders a random sphere scene with explicit quad + sphere lights for testing.

Features

GPU path tracer via Metal hardware ray tracing (Apple M-series)
CPU fallback with Odin thread pool
OBJ + MTL scene import with map_Kd texture loading (PNG, JPEG)
Scene graph with TRS transforms and instancing
Direct light sampling with multiple importance sampling
Explicit quad and sphere light primitives
5 material types: Lambertian, Metal, Dielectric, Principled, Emissive
Disney-style Principled BSDF (GGX + Fresnel + diffuse) with proper MIS for NEE on glossy surfaces
Roughness cutoff and glossy bias controls
Scene-scale biased GI defaults and quality presets
GI cache confidence rejection with diagnostic debug views
Firefly clamping
EXR writer (FP16, multi-layer AOVs) — .exr output

Requirements

Apple Silicon Mac (M1+)
Odin compiler (nightly)
macOS 14+ (for Metal ray tracing)