Bunny Flow Field

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. Goal and Visual Outcome

This project implements a particle flow-field engine. The core visual progression is:

• Particles start from a random chaos cloud;
• They evolve under a continuous dynamical system dx/dt = F(x);
• They gradually converge from disorder and attach to the Stanford Bunny surface;
• In the final segment, the camera performs a full orbit around the bunny to present a stable structure.

The target effect is a strong “chaos -> structure emergence” transition.

2. Overall Architecture

The engine pipeline has five stages:

1. Mesh loading
   Parse vertices and triangle faces from stanford-bunny.obj, then normalize scale and center the mesh.

2. Target distribution sampling
   Build a triangle-area CDF and uniformly sample points on the bunny surface as per-particle targets target[i].

3. Chaos particle initialization
   Initialize particle positions in a random spherical cloud with Gaussian noise; initial velocities are small random values.

4. Particle dynamics evolution
   Integrate particle position/velocity each frame. Early stage keeps chaotic perturbations; late stage switches to strong attraction and settling.

5. Projection rendering and output
   Project particles to 2D, accumulate energy, apply tone mapping, export PNG frames, then encode a GIF with ffmpeg.

3. Dynamical System Design

Each particle state is (p, v), with target point t.
The system uses two stage weights:

• alpha = morph_progress(t): structure formation progress;
• settle = settle_progress(t): late-stage strong convergence progress.

The force field is a blend of two components:

• F_chaos: sinusoidal noise + vortex + random jitter, creating early unordered motion;
• F_form: attraction toward target + mild spiral guidance + damping, producing stable geometric contours.

Overall form:

F = (1 - blend) * F_chaos + blend * F_form

blend increases with alpha/settle.
In the final stage, an additional snap term (position interpolation toward target) plus stronger damping is applied to make convergence more complete and sharper.

4. Convergence and Clarity Strategy

To avoid a blurry final frame, three key strategies are used:

• Dynamic trail decay
  decay decreases as settle increases, so historical trails are cleared faster in the final stage.

• Dynamic exposure
  Exposure is reduced near the end to suppress overexposed, washed-out contours.

• Outlier particle suppression
  Rendering intensity is weighted by convergence convergence; particles far from targets contribute less in the final stage.

5. Camera and Composition

The camera uses a segmented trajectory:

• First 70%: slow approach and observation to preserve continuity from chaos to structure;
• Last 30%: a full 360-degree orbit around the bunny to show the final form.

Field-of-view and orbit radius are tuned to avoid extreme close-up only in the last stage and improve overall readability.

6. Color System (Adjustable)

The engine supports three color parameter groups (0..1):

• --chaos-color r,g,b: main color in the chaos stage;
• --form-color r,g,b: main color in the converged stage;
• --bg-color r,g,b: background base color.

During rendering, color blending follows alpha + convergence + settle, creating a smooth transition from chaos color to structure color.