A stored drawing blended with random values at several noise levels. No diffusion model runs here, and the three panels are not a subtraction equation.
Compare noise levels
This illustration blends a stored drawing with random values. Compare the noise levels in either direction. It shows the idea of a noisy training input, not a trained model generating an image.
first viewnoise level 8
→
noise illustrationnot a prediction
→
comparisonnoise level 7
pure noisestored drawing
In a latent diffusion model, an autoencoder compresses an image before denoising. A decoder converts the resulting latent array back into pixels. This drawing-based illustration does not include either network.
Illustration only. Each frame is built directly from the known drawing and random values. The middle panel visualizes noise; it is not a learned estimate, and the right panel is not calculated by subtracting it from the left.