Noisy image

As explained in Build a Text-to-Image Generator (from Scratch)

A noisy image is an image formed by combining an underlying image with injected noise.

A noisy image combines an underlying clean image with injected noise. In forward diffusion, each time step produces a new noisy image as a weighted sum of the previous image and a normally distributed noise tensor with the same shape. section 1.3.1 section 5.1.1

A noisy image changes through the diffusion schedule:

• Start with the clean image x₀ at time step 0. section 1.3.1
• Inject a controlled amount of noise at each time step, producing x₁, x₂, and later noisy images. section 1.3.1 section 5.1.1
• At T = 1,000, the image approximates pure random noise; the time step records how much noise it contains. section 5.1.1
• Pass the noisy image and its time step to a denoising U-Net, which predicts the injected noise for reconstruction. section 5.2 section 5.3.1

Visualizing a noise schedule

Apply a DDIM noise scheduler to four clean flower images and display them at time steps 0, 200, 400, 600, 800, and 1,000. The images progressively lose visible structure until they become random noise. section 7.2.2

Figure 1.6 In the forward diffusion process, we start with a clean image from the training set, x₀, and add noise ε₀ to it to form a noisy image x₁, which is a weighted sum of x₀ and ε₀. We repeat this process for 1,000 time steps until the image x_1,000 becomes random noise.

Trace the forward process from a clean dog image x₀ through successive noisy images to random noise at time step 1,000. section 1.3.1

How Build a Text-to-Image Generator (from Scratch) covers it

How deep it goes

Explains weighted image-noise construction, arbitrary-time sampling, and U-Net noise prediction, but defers full reverse-process internals.

Examples

Uses dog images, clothing-item batches, and flower-image grids to show progressive corruption and denoising-model training.

What it assumes

Assumes clean image data and basic noise and time-step notation while defining the forward process from the clean-image starting point.

What you can do after

Apply it via a DDIM scheduler or diffusion training loop to create scheduled noisy images for denoising-model training.

How it connects

Connects noisy images to forward diffusion, reverse diffusion, diffusion models, denoising U-Nets, and noise schedulers.

liveBookmanning.com

Read Build a Text-to-Image Generator (from Scratch) on liveBook Get the book at manning.com

Related concepts

• U-Netused by · U-Net is an encoder-decoder neural-network architecture with downsampling and upsampling paths connected by skip connections for combining features at multiple spatial scales.
• noise scheduleproduced by · A noise schedule is a prescribed sequence of noise levels or coefficients governing noise addition and removal across diffusion timesteps.
• diffusion modelproduced by · A diffusion model is a generative model that learns to produce data by reversing a process that progressively adds noise to data.

This explainer was generated by Manning's knowledge graph from the text of Build a Text-to-Image Generator (from Scratch). The section links point to the parts of the book it draws on.