Bunny Dissolve

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. Project Objectives

This project implements a particle-based dissolve animation from stanford-bunny.obj. The core visual sequence is:

1. Build a static bunny silhouette
2. Trigger fracture from a single ear-tip seed (propagation-based)
3. Let particles disperse under normal-direction blast + Curl Noise field
4. Finish with either reconstruction (regroup) or disappearance (vanish)
5. Run a two-orbit camera path with near/far shot transitions
6. Export a PNG sequence and encode it into a high-quality GIF

2. Overall Technical Architecture

The system uses an offline rendering pipeline:

1. CPU side: OBJ parsing, triangle construction, area-weighted particle sampling
2. GPU side: per-frame particle dynamics update (CUDA kernel)
3. GPU side: particle projection + multi-segment trail splat accumulation (CUDA kernel)
4. GPU side: background composition, bloom, vignette, gamma (CUDA kernel)
5. CPU side: write per-frame PNG files
6. FFmpeg: generate GIF with palettegen + paletteuse

Key idea: instead of triangle rasterization, reconstruct the mesh surface as a dense particle cloud, then apply stage-based dynamics and stylized shading.

3. Geometry Data and Particle Initialization

3.1 OBJ Parsing and Triangulation

• Parse v and f lines, including first-field index parsing for f v/vt/vn format.
• Triangulate polygonal faces via fan triangulation.
• Skip degenerate triangles (very small normal length).

3.2 Normalization

• Compute the bounding-box center and translate geometry near the origin.
• Uniformly scale by the maximum extent to stabilize framing and motion scale.

3.3 Area-Uniform Surface Sampling

• Compute triangle areas and build a CDF.
• For each particle:
  • Pick a triangle with area-weighted probability
  • Use barycentric sampling $$b_0=1-\sqrt{u},b_1=\sqrt{u}(1-v),b_2=\sqrt{u}v$$
  • Position: $\mathbf{p}=b_0\mathbf{a}+b_1\mathbf{b}+b_2\mathbf{c}$
  • Store surface normal, random phase phase, and random seed seed

This yields a particle distribution that visually behaves like a true mesh surface instead of random volumetric sampling.

4. Procedural Noise and Curl Field

4.1 Hash-Noise Foundation

• Use pcg_hash + integer lattice hashing to construct parallel-evaluable pseudo-random value noise.
• value_noise_3d produces continuous noise via trilinear interpolation and Hermite smoothing.

4.2 Curl Noise

• Build a 3D vector potential field $\mathbf{F}(x)$, then compute curl: $${\mathbf{v}}_{curl}=\nabla \times \mathbf{F}$$
• Estimate derivatives with finite differences for stable and parallel-friendly evaluation.

Meaning: the curl field is approximately divergence-free, so particle motion looks like coherent advection/vortical flow instead of unstructured jitter.

5. Fracture Dynamics

Particles are updated each frame with explicit integration:

where d is drag and F is total force.

5.1 Ear-Tip Single-Seed Propagation Mask

Given the ear-tip source point ${\mathbf{s}}_{ear}$:

Interpretation: particles farther from the ear tip fracture later, producing a clear propagating wavefront.

5.2 Blast Pulse and Wavefront Enhancement

• A two-peak pulse controls primary and secondary bursts.
• source_gain strengthens the near-ear region.
• wave_front boosts the currently advancing fracture front.

Approximate blast term:

5.3 Force Composition

The total force is the sum of:

1. Normal-direction blast force (along surface normals)
2. Shock radial force (outward from ear tip)
3. Curl Noise flow force (smoke/fluid feel)
4. Global outward expansion force (detaching from body)
5. Tangential swirl force (adds rolling structure)
6. End-state force:
   • regroup: pull back to initial surface position
   • vanish: gravity-driven sinking + alpha fade

5.4 Stage-Based Opacity and Regroup Stability

• regroup: raise opacity after dispersion; final snap interpolation enforces stable silhouette closure.
• vanish: delay fade-out to preserve longer trail readability before disappearance.

6. Camera System (Two Orbits + Near/Far Shots)

The camera is a parameterized orbital camera:

• Total orbit angle: 4π (two full revolutions)
• Radius switching between far ~3.3 and near ~1.72
• Near-shot windows driven by two Gaussian pulses in mid/late timeline
• Mild vertical oscillation + target offset coupling

Effect: combines global shape readability (far shots) with high-impact fracture details (close shots).

7. Particle Rendering Model (Screen Space)

7.1 Projection

• Use camera basis vectors right/up/forward for view-space projection.
• Particle depth controls screen radius and base opacity.

7.2 Velocity Trails (Key Visual Enhancer)

• Project world-space velocity into screen space to get motion direction.
• Compute dynamic tail_len from speed magnitude.
• Use up to 4 trail_taps for reverse-direction multi-segment splats.
• Control each segment’s radius, alpha, and color independently (toward violet-pink at the tail).

This is the core implementation behind “longer pink-purple trails”, without relying on post-process motion blur.

7.3 Accumulation Strategy

• Accumulate all particles into an accum buffer with atomicAdd (RGB + alpha).
• Normalize color by density during composition.

8. Composition and Stylization

Implemented in compose_frame:

1. Deep-to-light purple gradient background
2. Dual-layer halo (central + offset halo)
3. Foreground-over-background blending (density-based)
4. Light bloom enhancement
5. Vignette for subject focus
6. Gamma correction (approximate sRGB)

Overall color strategy:

• Static pink-white c_static
• Mid-stage magenta c_mid
• High-energy deep purple c_deep
• Trail accent color c_tail

9. CUDA Parallelization and Memory Layout

9.1 Kernel Breakdown

• init_particles
• step_particles
• clear_accum
• raster_particles
• compose_frame

9.2 Data Layout

• Particle state: float4 arrays (initial, position, velocity)
• Accumulation buffer: float4 (RGB sum + alpha sum)
• Output frame: uchar3 (stored as linear byte array)

9.3 Complexity

• Simulation: $O(N)$
• Rasterization: approximately $O(N\cdot r^2\cdot taps)$, affected by particle radius and trail segments
• Main bottleneck: atomicAdd contention in high-density regions