Diffusion models learn to reverse a gradual noising process, denoising a random point cloud step by step until a structured shape emerges. GANs instead pit a generator against a discriminator in an adversarial push/pull that sharpens samples without an explicit step schedule.
Diffusion: x_{t-1} = x_t - eps_theta(x_t, t) * dt + noise
GAN: G tries to fool D; D tries to catch G
- Mode — switch between diffusion-style iterative denoising and GAN-style adversarial jitter.
- Denoising steps — number of iterations to morph noise into the target shape (diffusion mode).
- Noise level — initial variance of the point cloud (how "noisy" it starts).
- Guidance strength — how strongly the process is pulled toward the target structure each step.
Watch the particle field settle from a random cloud into a torus knot shape as the sampler runs — this is a visual metaphor for the generative sampling loop, not the literal pixel space of a real diffusion model.