DALL-E and Stable Diffusion both use latent spaces — your text prompt gets encoded to a point (or region) in latent space, and the decoder turns that point into an image. Dragging through latent space = changing the generated image.
What happens to the decoded output as you move between concept regions?
This lab builds on concepts from:
Variational Autoencoders
2D Latent Space — drag to explore
Decoded Output
Neutral
Decoded: A neutral expression — flat mouth, relaxed brow, open eyes at rest
Key insight: A VAE encodes each input not as a single point but as a distribution (μ, σ). During generation, we sample from that distribution — giving VAEs their generative capability. The smooth blending between regions shows the continuous, structured latent space.
Reparameterization: z = μ + σ·ε
where ε ~ N(0,1)