Wood Bunny

woodbunny

The source cuda code can be found here, which you should noted is that the stb_image.h, stb_image_write.h and the stanford-bunny.obj must be placed at the same file level.

1. Basic Rendering Equation

For any surface point $x$ and view direction $ω_{o}$ , we use the standard surface rendering equation:

L_{o} (x, ω_{o}) = L_{e} (x, ω_{o}) + \int_{H^{2} (n)} f_{s} (x, ω_{i}, ω_{o}) L_{i} (x, ω_{i}) ∣ n \cdot ω_{i} ∣ d ω_{i}

Where:

$L_{o}$ : outgoing radiance
$L_{e}$ : emitted radiance (only for light sources)
$f_{s}$ : surface BSDF (both wood and ground are non-delta)
$L_{i}$ : incident radiance

In this implementation, both the bunny and the ground are non-specular-delta materials, so both can use NEE + MIS.

2. Wood Material Model

woodbunny.cu uses an empirical model: “procedural texture + diffuse + clear-coat-like specular”:

f_{w oo d} = k_{d} \cdot \frac{ρ _{w oo d} ( x )}{π} + k_{s} \cdot f_{s p ec}

And:

k_{d} = 1 - k_{s}, k_{s} = bunny_spec_strength

The specular term $f_{s p ec}$ uses a power-cosine (Phong-like) form, with exponent mapped from roughness.

3. Procedural Wood Grain $ρ_{wood} (x)$

3.1 Multi-frequency Noise (FBM)

FBM is built by summing value noise:

fbm (p) = o = 0 \sum N - 1 a_{o} n (2^{o} p)

Used for:

Coordinate warping
Long-direction grain details
Pore darkening mask

3.2 Ring and Grain Fusion

The main wood pattern comes from “radial rings + perturbation”:

r = x_{w}^{2} + z_{w}^{2}, ring = 0.5 + 0.5 sin (2 π \cdot (r \cdot s_{r in g} + w \cdot a_{r in g}))

Then it is combined with longitudinal grain and streak into a blend factor $t$ :

t = 0.58 ring + 0.32 grain + 0.10 streak

The final color is obtained by interpolating between light and dark tones:

ρ_{w oo d} (x) = lerp (C_{d a r k}, C_{l i g h t}, t)

And pore darkening is applied:

ρ_{w oo d} \leftarrow ρ_{w oo d} \cdot (1 - s_{p ore} \cdot m_{p ore})

4. Direct Lighting Estimation: NEE

For the rectangular area light $A$ , next-event estimation is used:

L_{d i rec t} \approx \frac{f _{s} ( x , ω _{i} , ω _{o} ) L _{e} ( y , - ω _{i} ) ∣ n _{x} \cdot ω _{i} ∣ ∣ n _{y} \cdot ( - ω _{i} ) ∣}{∥ x - y ∥ ^{2} p _{A} ( y )} \cdot V (x, y)

Where $V (x, y)$ is visibility evaluated by a shadow ray.

5. MIS (Light Sampling + BSDF Sampling)

The implementation uses the power heuristic:

w_{a} = \frac{p _{a}^{2}}{p _{a}^{2} + p _{b}^{2}}

Multiply light-sampling contribution by $w_{light}$
Multiply BSDF-path-hit-light contribution by $w_{bsdf}$

This is more stable under small area lights + rough highlights, and significantly less noisy than a single strategy.

6. Path Throughput Update

At each bounce:

β \leftarrow β \cdot \frac{f _{s} ∣ n \cdot ω _{i} ∣}{p ( ω _{i} )}

After a certain depth, apply Russian Roulette:

p_{s u r v i v e} = min (max (β), 0.98)

This keeps the estimator unbiased while controlling computation cost.

7. Firefly Suppression and Denoising

7.1 Firefly Clamp

Apply soft clamping to overly bright sample radiance to suppress high-frequency bright speckles:

L \leftarrow L \cdot \frac{L _{ma x}}{max ( L )} (max (L) > L_{ma x})

7.2 Post-process Denoising (HDR Joint Bilateral)

In tonemap_ui_kernel, perform HDR-domain joint bilateral filtering before tonemap/gamma:

Spatial weight: based on neighborhood distance
Color weight: based on HDR difference and adaptive to spp

This significantly reduces noise while preserving wood-grain boundaries.