Final Lab (optional): Stable Diffusion

Open this notebook. Lab machines will work best!

• Stable Diffusion Deep Dive (name: u13n2-stable-diffusion.ipynb; show preview, open in Colab)

Try the following:

• Change the random seed in the main loop to several different values. What happens to the images?
• Change the prompt. Try more or less descriptive prompts.
• Try changing the guidance scale to 0. Generate a few different images. Does it seem to be using the prompt at all? Follow the flow of data through the code to see what’s happening.
• Try stopping the diffusion process early (if i == 10: break or the like in the loop). What do you notice about the output?
• Try starting the image2image model with latents that have a different amount of noise (change start_step; refer to num_inferenece_steps for its max value).
• Under “messing with embeddings”, try changing the two prompts to different texts. Or change the mix factor. What happens?