The hard problems in agent-made motion graphics.

An agent can write a page of HTML in a minute. Turning that page into a film that looks designed is the hard part. These are the 13 problems vawe works on, with the measured result for each. 7 are solved. 6 are still open. The acceptance table has 17 rows: 16 are measured by code and 1 by the judge.

The plan for the open problems is on the roadmap. Numbers come from vawe's own runs and tests. vawe runs on your machine, needs no account and has no per-render fee.

Solved

Each of these has a mechanism in the engine and a measured result.

01Solved
A browser is not a video renderer
Why it is hard
A web page reads the clock, runs timers and calls Math.random. Two renders of the same page give two different films.
Where vawe stands
A virtual clock owns Date, requestAnimationFrame, timers and Math.random. The renderer seeks CSS and Web Animations to any fractional time. A script error in the page stops the render. Tests check that frames are identical.
1 clockowns Date, rAF, timers and Math.random
02Solved
No private format
Why it is hard
An agent writes best what it already knows. A private format is one more thing to learn, and agents get it wrong more often.
Where vawe stands
A film is one HTML page. vawe had its own JSON engine, ran an A/B test, and removed the engine. Agents now write HTML, CSS and Web Animations.
```
<meta name="duration" content="6">
<h1>2.0</h1>
<style>
  h1 { animation: in .6s both }
</style>
```
1 fileone HTML page is the whole film
03Solved
Motion blur and finish
Why it is hard
A browser draws sharp frames, so fast motion looks like a slide show. Smooth gradients show bands, and the files are large.
Where vawe stands
Adaptive motion blur uses up to 32 subframes, set from the measured travel of each edge. It covers clips and rotation. Dither removes banding. Final files are smaller, and a short AV1 copy serves the web.
35 to 17 MBper 5 s final, plus a 2 to 4 MB AV1 web copy
04Solved
Designed motion
Why it is hard
Browser eases suit interface changes. Film motion needs the speed and influence handles that motion designers set in After Effects.
Where vawe stands
vawe uses After Effects speed and influence eases, fitted to real keyframes. All 94 moves use them.
1,339real keyframe segments used for the fit
05Solved
A weaker model can do the work
Why it is hard
A framework that needs the strongest model costs more per film and fails when the budget is small.
Where vawe stands
On a 22 s film, Sonnet matched Opus: the best draft scored 52 of 70 for both. Sonnet also raised its measured acceptance table from 5 to 15 of 17 rows.
52 of 70best draft score, Sonnet and Opus
06Solved
Fast iteration
Why it is hard
Each draft costs time, and the agent must read what the draft prints. Slow drafts and long logs both limit how many tries a film gets.
Where vawe stands
A 5 s draft takes 20 s, down from 27 s. Draft output is 4 lines, down from 28.
27 s to 20 sfor a 5 s draft; output 28 lines to 4
07Solved
Learning from traces
Why it is hard
Agents fail in ways the author does not predict. The failures show only in what the agents did.
Where vawe stands
vawe studied 147 agent transcripts. In them, 443 of 5,520 shell calls had failed. The causes were fixed.
443 of 5,520shell calls failed in 147 transcripts

Still solving

These are open. Each entry says what is known and what is not.

08In progress
Every AI film looks the same
Why it is hard
Models pick the same first idea, so their films share one brand, one shape and one subject.
Where vawe stands
Of 8 early films, 4 named the brand "Vesper" and all 8 had a glowing circle. An attractor check fixed those. Then all 3 complete films chose a finance tool. The convergence moves to the next attractor.
4 of 8early films named the brand "Vesper"
09In progress
Judging taste with an AI
Why it is hard
An AI judge scores the same film differently on different runs. A score that moves cannot guide a fix.
Where vawe stands
The judge varies by about 1 point per axis. A ledger, settled frames and pixel sizes reduced that. No complete film has passed every axis at 8 or more yet. A pairwise judge is next.
about 1 pointjudge variation per axis
10In progress
Making good measurable
Why it is hard
A check can only enforce what it can measure. Much of what makes a film good has no number yet.
Where vawe stands
The acceptance table has 17 rows. Code measures 16 of them and the judge scores 1. One idea, the eye path and a thread through the film are not measured yet.
16 of 17acceptance rows measured by code, 1 by the judge
11In progress
Final render speed
Why it is hard
A final render captures every frame at full size and rate, and capture is the slow part.
Where vawe stands
A 22 s film takes about 28 minutes to render. Most of that time goes to screenshots.
28 minfor a 22 s final, mostly screenshots
12In progress
Reliability
Why it is hard
A long render can crash late, and an agent that runs unattended does not see it.
Where vawe stands
A final render crashed at 96 percent and nobody saw it. A fix that resumes the render and reports the crash is next.
96%where a final render crashed unseen
13Parked
Sound
Why it is hard
A synthesized cue sounds machine-made. Good sound needs a better source than synthesis.
Where vawe stands
Sound work is parked. The synthesized cues sound machine-made, so vawe does not claim sound as a strength.
parkedsynthesized cues sound machine-made