L25 | English F3 Student Learning Log

Persona: I'm a Form 3 (Grade 9) student and I only read English. I know basic Python (variables, for loops, lists) and school algebra up to quadratics. I learn from the official pages, the videos with captions, and the book. Kaggle notebook steps are written as a practice log that follows the official notebooks.

What I understood (in my own words)

Sound can be drawn as a picture (a spectrogram), so image models can generate sound.
Diffusion on big images is slow; compressing to latents first is much faster.
The 'Latent' in Latent Diffusion means doing diffusion in the compressed space.
Everything from Part 2 connects together in this lesson.

Where I got stuck, and how I got unstuck

Where I got stuck	How I solved it
Reading a spectrogram	I whistled into a spectrogram app on my phone and saw a bright line that moved up as the pitch went up.
I only half-understand a lot of Part 2	I lined up all 17 concept maps so I know where each lesson fits in the whole pipeline. I'll come back for details when my maths is stronger.

My answers to the exercises

Q1. Stable Diffusion's VAE compresses a 512×512×3 image into 64×64×4 latents. By what factor?

Original: 512 × 512 × 3 = 786,432 numbers. Latents: 64 × 64 × 4 = 16,384 numbers. 786,432 ÷ 16,384 = 48 times.

Q2. Why does turning sound into a spectrogram let an image diffusion model generate sound?

A spectrogram turns sound into a 'time × frequency' image where brightness is loudness. A diffusion model only understands images, so it can generate a new spectrogram, which is then converted back into sound.

Q3. Looking back at Parts 1 and 2, write the three concepts you think matter most and explain how they connect.

Gradient descent: every model learns with it. 2) Loss functions: they set the direction of learning (MSE blurs, perceptual loss sharpens). 3) Pretraining and transfer: from resnet18 in L01 to the VAE in L25, we build on models others trained. How they connect: the loss defines the goal, gradient descent moves towards it, and pretraining gives a better starting point.

My Zettelkasten card

Card｜Change the representation and old tools solve new problems

The final lesson taught me a big idea. Sound isn't an image, but drawn as a spectrogram, an image diffusion model can generate birdsong. Images are too big to process, so compress them into latents first. When a problem is stuck, the answer is often not a new tool but a new way of representing the problem.

It reminds me of analysing my football matches. Watching videos, I couldn't say clearly what I was doing wrong. Then my coach had me mark where every one of my passes went on a pitch diagram. After one match I saw that most of my passes went sideways and almost none went forward. Same match, but drawn as a chart, the problem was obvious. A new representation also loses information (the diagram doesn't show timing), so you have to choose one that fits the question.

My next step: record ten seconds of a ball being kicked and headed, and compare their spectrograms in a phone app.

🔗 Related notes to grow next: → Problem solving: 'A new representation can make a hard problem easy' → Sport analysis: 'Chart the match and weaknesses appear' → Reflection: 'The map of the whole course matters more than any single detail'

Back to lesson note：L25_潛在擴散_202611041526