How Image Generators Work by Hand: Noise Schedules, Forward Noising, Denoising Steps, Guidance Scale & Latent Sizes
Follow a toy diffusion model through its numbers: noise schedules, how an 8-pixel image is noised, signal-to-noise per step, one denoising step, guidance scale and why latent images are cheaper, with every cell checked as you go.
One-time payment · lifetime access · lessons, labs, exam and certificate included · no subscription.
About this course
This course takes a toy image generator apart, one cell at a time. You scale an 8-pixel greyscale strip to −1…1 and back, build linear and cosine noise schedules with the running product ᾱ, and noise the strip to several steps with one fixed noise column. You turn the schedules into signal-to-noise ratios in decibels, rebuild an image estimate from predicted noise and take one reverse step. You mix prompt and no-prompt predictions with a guidance scale, price step counts in seconds and cost per 1,000 images, and compare pixel and latent tensor sizes and memory.
Lessons
- A Toy Image in Numbers: Pixel Values, Scaling to −1…1 and Back Pro
- Noise Schedules: Linear vs Cosine β and the Running Product ᾱ Pro
- Forward Noising: Mixing Image and Noise at t = 0, 5, 10 and 20 Pro
- Signal-to-Noise: SNR per Step, Decibels and Where the Image Disappears Pro
- Denoising: From Predicted Noise to an Estimated Image and One Reverse Step Pro
- Guidance: Mixing Prompt and No-Prompt Predictions with a Guidance Scale Pro
- Steps vs Time: Seconds per Image, Cost per 1,000 Images & Diminishing Returns Pro
- Latent Space: Pixel vs Latent Tensor Sizes, Compression Factor & Memory per Batch Pro