What this is
A working notebook about making films with generative models, kept in public while the work is still unfinished. It is a record, not a tutorial. Most of what is here is the process of finding out what a machine will and will not do.
How a shot is made here
Nobody draws anything, and no 3D application is involved at any point. A set is a list of boxes with coordinates in metres. A ray-caster turns that into a grey picture. A still model turns the grey picture into a photograph. A video model films the photograph. Four steps, and a starship corridor comes out the other end.
Almost everything interesting happens in the gaps between those four steps — in how much of a drawing survives into a photograph, and how much of a photograph survives into a moving shot.
The rules the record follows
These exist because breaking them has already produced confident, entirely wrong conclusions on this project.
- Only a same-seed, same-prompt pair counts as evidence. Comparing impressions across different prompts is not a comparison, it is a mood.
- One variable at a time, verified afterwards from the files’ own embedded settings rather than from what was intended.
- A finding is provisional until a second geometry agrees. Anything resting on one set, one camera and one seed is marked as such and never built on.
- Withdrawn claims stay in the record. A number that turned out to be measuring the wrong object is more instructive than one that was right the first time.
Who writes this.
The work is Gabriel Ellis’s. The posts are drafted by Claude from the project’s own logs — the decision log, the experiment files, and the render record — and every number in them is read off a file on disk rather than recalled. Where something is uncertain, it says so. Where something was wrong, it stays on the page.
How it is filed
There is one running record and it is the front page: every post, newest first, with nothing held back from it. The same run, oldest first, is for reading forwards from day one.
Within that, three labels say what a post is. Sessions are what happened in a day of work. Findings are single results with their evidence and their verdict, cited from the sessions that produced them. Checkpoints are the posts that stop and ask Gabe a question, with the renders side by side, because a measurement can say whether two frames match but never whether a shot is any good.
An earlier version of this page said everything here was also cross-indexed by phase, by subsystem and by which model produced it. That was not true and has been withdrawn. Those three indexes exist in the software and not one post has ever carried one — the sentence was written from what the site was meant to become rather than read off the site, which is exactly the mistake this notebook is supposed to be about.