How Image Generators Work by Hand: Noise Schedules, Forward Noising, Denoising Steps, Guidance Scale & Latent Sizes

What you'll build

What you will be able to do

Each lab ends with a concrete result that you build yourself and the checker confirms, cell by cell.

• Turn an 8-pixel greyscale strip into the −1…1 numbers an image generator works with, measure its contrast, and turn a model output back into valid pixel values with clipping.
• Build two noise schedules for a 20-step toy run: a linear β with its running product ᾱ, and a cosine ᾱ with the β that belongs to it.
• Noise the Mistlet strip to four steps of the schedule with one fixed noise column and measure how far the pixels move away from the clean image.
• Turn both schedules into a signal-to-noise ratio in decibels, find the step where noise becomes louder than the image and chart how the two schedules fade.
• Take the network's noise guess at t = 10, rebuild an estimate of the clean image, measure its error and take one reverse step back to t = 9.
• Combine two noise predictions with a guidance scale, see how the estimated image changes and find the scale from which pixels are pushed out of range.
• Price six step counts for the Mistlet sampler: seconds per image with guidance, cost per 1,000 images, extra quality per extra second and the cheapest setting that reaches a quality target.
• Count the numbers in pixel and latent tensors for four image sizes, work out the compression factor and compare the memory a batch needs.