ContentsThe library

How Noise Becomes a Picture

The Same Model, Twenty Times Faster

Last timeSteering the Walk

Removing the added noise turns the sampler into a fixed path that can be travelled in any number of steps, which is where the twentyfold speedup came from.

The sampler of the earlier lesson takes a thousand steps because the forward

process had a thousand steps. Nothing about generation requires that number, and

this lesson shows where it goes. Nothing is retrained; the weights from lesson

three are used exactly as they are.

Removing the noise, properly this time

An earlier lesson established that dropping the added noise from the reverse step

collapses the output to a single smooth composite. The deterministic sampler is

not that. The reverse step was derived to undo one forward step, and its centre

is only meaningful one step at a time. The deterministic update is built

differently: it estimates the clean picture from the current position and the

predicted noise, then re-noises that estimate to the level of the target time.

FIG 1A deterministic step from time t to any earlier time s
any earlier time, not necessarily the one immediately before t
the implied guess at the finished picture, from lesson three
the signal and noise factors belonging to the target time
Read it as a round trip: jump all the way to a guess at the finished picture, then travel back out to the noise level of the target time using the same predicted noise. Because both ends of the trip are written in terms of the closed form, the target time can be anywhere, and nothing is drawn at random.

The step list is a choice

The thousand times of training now function as a menu rather than an itinerary.

Choose fifty of them and run the same loop. Nothing in the trained network knows

or cares how many entries the list has, because the time is an input and every

one of those times was trained on.

The lesson stops here

3 more paragraphs to go

You have read the opening. The rest of the argument, the problems that check whether it landed, and the lines worth keeping at the end all come with a plan.

The first lesson of every course in the library reads the whole way through, free, so you can see exactly what the rest of them are.

See the planThe contents

This is the reading half

Starting the course gives you your own copy of it. Every idea on every page has problems standing under it, marked with a reason rather than a tick, and any sentence you do not believe can be opened and argued with. None of that can happen on a page nobody owns.

The contents

The rest of this course

  1. 01A Thousand Small Steps From a Photograph to Static
  2. 02A Thousand Steps Collapse Into One Lineopening only
  3. 03Five Lines of Training, and One Squared Erroropening only
  4. 04One Step Back, a Thousand Timesopening only
  5. 05The Network Is Pointing Uphillopening only
  6. 06Overshooting the Condition on Purposeopening only
  7. 07The Same Model, Twenty Times Fasteryou are here
  8. 08If the Path Is Yours to Pick, Pick a Lineopening only

Read alongside