Session 1 · Tool Lab · Lesson 02
How Machines Read Images
Pattern extraction, semantic clustering, attention weighting, latent space mapping. No understanding — only prediction.
Concept
- A model reads an image as patterns and probabilities, not as a scene.
- Attention weighting decides what counts; everything else is discarded.
- Generation is denoising in reverse — training and inference share the same text-conditioning machinery.
If the model has no understanding, only prediction — what is it giving you back?
Slides
- 07
How AI “Sees” Images
- Pattern extraction
- Semantic clustering
- Attention weighting
- Latent space mapping
- No understanding, only prediction
- 08
Text-Guided Diffusion: How Training Becomes Inference
- Training — learn text-conditioned denoising
- Learns to predict noise at many noise levels, conditioned on text
- Training produces the weights inference uses
- Inference — generate from text with fixed weights (no updates)
- Same text-embedding interface, replayed as iterative sampling from noise
From the deck · slide 08
Student activity
15:00
- 01Ask ChatGPT to describe one image, and save its description verbatim — including its mistakes.
ChatGPT
Multimodal chat with a free tier. The workshop's default interpreter — give it an image and it will describe, read, and re-read what it sees.
Upload your image, send the prompt below, and copy the reply into Machine description — verbatim.
Prompt to try
“Describe this image as literally as you can. List the objects, people, and setting you see, and say what is happening — no interpretation.”
External tool — it has its own privacy policy and may change or require an account.