Every deep generative model turns a random point — noise — into a structured sample that looks like it came from the training data. This scene shows that transformation as a cloud of particles moving from a scattered gray blob (pure noise, left) to a colored knotted curve (the learned data manifold, right). The shape of the motion is different for each architecture, and that shape is the whole point.
z is first sampled from a small Gaussian "bubble" (reparameterization trick), then decoded smoothly into data space. Because the decoder is trained to maximize a reconstruction likelihood under a simple output distribution, the result keeps a small residual blur — VAE samples are famously a bit softer/blurrier than GAN or diffusion samples.Diffusion models such as DDPM and score-based generative models can need dozens to thousands of denoising steps at generation time, which is why so much research effort (DDIM, distillation, consistency models) has gone into cutting that step count while keeping GAN-like sample quality.
A particle cloud starts as pure noise in latent space and morphs into a structured "data manifold" shape — watch how a GAN's single forward pass, a VAE's sample-then-decode step and a diffusion model's many small denoising steps each get there differently.
All three families map noise to data, but the path differs: GANs use one deterministic non-linear warp, VAEs sample a Gaussian latent code before decoding (leaving mild residual blur), and diffusion models take many small denoising steps that visibly settle in stages.
Pick a model from the dropdown, adjust diffusion step count and playback speed, and watch the progress bar and stats track the phase, step and remaining noise level. Click "New sample" any time, or leave looping on to keep regenerating automatically.
Reducing the number of diffusion sampling steps while keeping quality — via methods like DDIM, distillation and consistency models — has been one of the most active research areas in generative modeling since 2020.