Lesson 10

Puru - a bunch of images of (spherical) linear interpolation between two different starting points

Student examples of making different pictures

• Going from an old car prompt to a new car prompt
• dinosaur -> to a bird
• dog -> unicorn
• classic optimizer efforts, couple of hours

Recap

• With Handwritten digits,
• start with a 7, add some noise and get a "noisy 7"
• and feed this into Unet, and try to predict the noise
• we can also pass in the actual number 7 as "guidance" to help it along
• Turn captions into embeddings, via using existing images + captions (alt tab)
• Then train a text encoder + image encoder
• compare the dot product
• contrastive loss
• take the text encoder and feed into

--image--

Doing inference

• put a prompt + some noise
• Unet will predict noise
• will subtract gradually, and then resubmit the updated photo
• used to take 1000 steps, now takes 60 (maybe less!)

--image--

Papers

A lot of interest + work still being focused on this work.

Progressive Distillation for Fast Sampling of Diffusion Models On Distillation of guided Diffusion Models Imagic: Text-Based Real Image Editing with Diffusion Models

Overview of "Progressive Distillation for Fast Sampling of Diffusion Models"

Hand-written Notes