Where the Machine Is and Isn't

The short version is on the about page and it is true but compressed. This is the uncompressed one, including the places we tried something and stopped.

The line, in ten steps

Every episode goes through the same sequence, numbered here so we can point at the numbers later.

1. 1

   The mood — no tools

   One word, chosen by a person from a long list a person keeps. Rain. Hearth. Harvest. We have never asked a model for a mood and do not intend to. This step decides what the episode is for.

2. 2

   Reference collection — no tools

   Photographs of real places — real cottages, real copper pots, real light at the real hour. Gathered and argued over by hand; roughly half the schedule lives here. Nothing is generated, because the whole point is to look at things that exist.

3. 3

   Scene design — no tools

   Floor plan, weather schedule, timeline of small events. Where the door is, where the light comes from, when the kettle goes on. Three ordinary documents, written by people.

4. 4

   Palette — no tools, and this is deliberate

   Pulled from one real photographed object. We tried asking a model for palettes; the results were pleasant and characterless in a way that took a while to name — the average of every cosy interior ever posted, which is exactly what we are trying not to be.

5. 5

   Key frames — this is where the machine is

   The stills. Composition, staging, light sources and palette are fixed beforehand and held; the pipeline serves them. Most of what comes back is discarded, and what survives is chosen by a person against the reference board from step two.

6. 6

   The hand pass — people, on everything that matters

   Hands, faces, the fall of light on any surface the eye returns to. If a viewer would look at it twice, a person has been over it. Slow, unglamorous, and the difference between our frames and the same pipeline used carelessly.

7. 7

   Motion — mixed, and supervised closely

   Layers move; some of that is assisted, all of it is timed and corrected by a person. Timing is the thing we will not hand over — reasons below.

8. 8

   Sound — people, and mostly a kitchen

   Recorded, layered and mixed by hand; much of the library was made with a bucket, a colander and a lot of patience. No generated audio in any episode, and no generated music, because we do not use music.

9. 9

   Assembly — people, with ordinary software

   Loop structure, offsets, order of events, where the episode ends. Ordinary editing tools. Nothing interesting happens here except decisions.

10. 10

    Titles, covers, descriptions, translation — mixed, and labelled

    Titles and descriptions are written by a person in English. This site exists in twenty languages and we did not write twenty: a model does the first pass and we correct what we can read. Where the words are a viewer's own, the translation is machine-made, marked as such, original underneath.

So: steps five, seven and ten. Not the whole line, and not none of it.

What we tried and stopped doing

Probably more useful than the current state: the things we backed out of.

Generated music. Three episodes, early on. Competent, and it flattened every scene it touched — and it solved a problem we did not have, since the episodes are better with no music. Dropped, and the idea of music with it.

Generated descriptions. About six weeks. They were fine, they read like every other description on the platform, and one described a scene that was not in the episode. That last part is the reason: a promise about the contents that nobody checked is not a promise.

Upscaling the old episodes. Eleven early files at a lower resolution, and an obvious fix. The results were sharper and worse — fire came back with a plastic edge, and grain we put there on purpose was read as noise and removed. Abandoned.

Anything touching timing. Tested twice, on the placement of small events. Both times the result was regular in a way that felt like a metronome.

Everything we have handed over and taken back, we took back for the same reason: it produced the average of the genre, and the average of the genre is what we are trying to get out of.

Where a person is not replaceable

Two places, both about judgement rather than craft.

The first is timing. When the rain doubles, how long the cat waits before giving up on the moth, how many seconds of nothing fit between two small events before the room goes slack. We cannot describe this well enough to delegate it, and every attempt to systematise it produced something that felt scheduled. Evenings are not scheduled; the illusion depends on events arriving at intervals nobody could have predicted and everybody accepts afterwards.

The second is deciding what not to include. Almost every improvement in our work has been a subtraction: the second saturated colour removed, the extra moving element cut, the clever shot dropped for being clever. A tool will happily give you more. Nothing gives you less. Somebody has to say that the episode is finished and the good idea stays out.

The reference question

This gets its own section because it is what people are usually actually asking about.

We do not use living artists' names as instructions, and we do not use a studio's name as a style token. References are photographs of real places; the look is built from palettes pulled off real objects. When a frame starts to resemble a specific work rather than a real thing, it goes back on the bench — because a room that reminds you of a film is a room you are looking at rather than sitting in.

And then there is our name. Ghibli Mood ON wears an inspiration on its front, and that is not a neutral choice. It is a debt, stated openly in the first two words anyone reads. What we can say is what it is not: not an attempt to be mistaken for that work, no claimed connection, and nothing in the pipeline aimed at reproducing anyone's drawings. What we borrowed is a temperature — attention paid to small domestic things, weather treated as a character, quiet allowed to last. Whether the name is the right way to signal that is an argument we have had more than once and have not closed.

Why we publish this at all

Nobody requires it. Plenty of channels here use the same tools and say nothing, at no cost, and some of them do fine work.

We publish it for one reason: a person deciding whether to spend an evening inside something we made has a right to know what made it. Not because tools are shameful, and not for credit — simply because withholding it would be a decision to make people ask.

And this will probably change. The tools move, our opinions move, and there are at least two steps on that list we have argued about this year.

So, plainly: if our position changes, we will update this page and say what changed and when. We will not edit it quietly. A version of this piece with a paragraph silently removed would be worse than never having written it — it would turn a statement into a decoration.

The part where we don't defend it

This piece will not satisfy everyone, and we are not going to write a closing paragraph designed to bring you round.

Some readers hold that any generated frame disqualifies the work. That position is coherent, we are not going to argue you out of it, and this page exists so you can decide without having to ask. Others think the disclosure is unnecessary fuss. We would rather be tedious about it than smooth.

Underneath all of it is a room somebody spent six weeks getting right, meant to be useful to a person at eleven at night. The rest of how that gets built is in how an episode is made.

The evening is warm here.