Stable Diffusion: from Practice to Understanding from the Inside (fast.ai)

Topics covered

• •Run Stable Diffusion through the `diffusers` library in Colab or locally.
• •Master the core generation controls: prompt, negative prompt, guidance scale, number of steps, seed.
• •Try img2img mode and understand how textual inversion differs from Dreambooth.

Topics covered

• •Assemble the Stable Diffusion pipeline by hand from four separately loaded components.
• •Understand the role of each block: the CLIP text encoder, the VAE autoencoder, UNet, the scheduler.
• •Trace the path of text from a string to embeddings and learn to intervene in it.

Topics covered

• •Understand the conceptual basis of diffusion models without relying on ready-made libraries.
• •Connect "noise prediction" to the idea of a gradient over the data (the score function).
• •Understand how the training process is structured: noising, the objective, the noise schedule.

Topics covered

• •Write a sampling loop from scratch, without relying on ready-made schedulers.
• •Understand the differences between sampler families (DDPM, DDIM, Euler, ancestral variants) at the level of ideas.
• •Learn to experiment: measure the effect of changing the noise schedule, number of steps, and guidance.

Topics covered

• •Combine all the course's skills into a completed independent piece of work.
• •Demonstrate an understanding of SD's architecture at the level of "I can modify it", not just "I can run it".
• •Present the result in a form suitable for a portfolio.