L22 | 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)

A sampler decides the route from noise to image; different samplers trade speed against quality.
Euler is the simplest: look at the current direction and take one step.
Heun is more careful: take a trial step, check the direction there, then average the two to correct the step.
Scaling numbers to standard deviation 1 helps the model learn, just like in L17.

Where I got stuck, and how I got unstuck

Where I got stuck	How I solved it
The paper's equations	I focused on Jeremy's diagrams and code to see how each sampler takes a step.
What is σ (sigma)?	The size of the noise. The bigger σ is, the noisier the image.

My answers to the exercises

Q1. What's the difference between Euler's and Heun's method? Use walking as an analogy.

Euler is like checking your compass once and walking a step. Heun is like taking a trial step, checking the compass again at the new spot, averaging both directions, and redoing the step. Heun checks twice per step, so it's slower but more accurate.

Q2. Why scale the model's inputs and outputs to standard deviation 1? (Hint: link to L17.)

As in L17, numbers that are too big or too small explode or vanish as they pass through the network, making learning hard. Scaling to standard deviation 1 keeps inputs at every noise level in a similar range, so the model learns more easily.

Q3. Ancestral Euler adds a little random noise back at each step. How might that affect the variety of generated images?

Adding a bit of randomness each step means the same start can follow different routes, so images are more varied. Too much randomness could make images unstable or blurry.

My Zettelkasten card

Card｜Testing the ground before committing makes each step more accurate

Heun's method is like a careful hiker: take a trial step in the current direction, check the direction from there, average the two, then take the real step. It costs an extra calculation, but each step is more accurate, so you can need fewer steps overall.

I've found the same in maths homework. I used to write the answer as soon as I'd calculated it and lost marks on small slips. Now I substitute the answer back into the original equation before moving on, and fix it if it doesn't fit. Each question takes 30 seconds longer, but I make far fewer mistakes and save time overall. Not every question needs this, though. For easy ones I just go, like how Euler is good enough in some cases.

My next step: check every maths homework answer by substitution and count how many mistakes it catches.

🔗 Related notes to grow next: → Study habits: 'Checking once is cheaper than redoing' → Numerical methods: 'Predict-then-correct appears in many algorithms' → Research: 'Read the diagrams and code before the equations'

Back to lesson note：L22_Karras論文_202611041526