nolanlabs

Blog · Agents

Titles and motion graphics with an agent

· Updated · 6 min read

A long red cloth hangs in midair, moving, over a flat and empty pale plain.

Short answer

An agent writes a small web page, and a renderer draws it one frame at a time, so words stay editable and a frame comes out the same each time. Remotion does this in React and HyperFrames in plain HTML. Generated video makes pixels instead, which suits footage, not type.

On this page
  1. Why draw a page one frame at a time
  2. A title, worked through
  3. How it differs from Remotion and HyperFrames
  4. How it differs from generated video
  5. Where this falls short
  6. How it works in nolanlabs

The page is HTML for the words, CSS for the type and a few lines of animation. The renderer opens it in a browser with no real clock, sets the time to exactly one frame, takes a picture, and repeats.

Facts about other products below were read from their docs on November 4, 2026. One of the systems compared is ours, and we say which.

Why draw a page one frame at a time

Because it turns time into a number. A browser normally plays animation against the clock, so a slow machine skips moments. A renderer instead asks what frame 90 looks like, and works out the time as 90 divided by the frame rate. HyperFrames’ determinism page lays out this model well: render time never reads the real clock, so a frame that takes a long time to draw still lands in the right place.

For an agent, the appeal is practical. It already writes web pages. Text, layout, charts and curves are things it can read back and change, and it can look at a rendered frame to check itself.

A title, worked through

Here’s a complete page for a short title. It’s the sort of thing an agent writes, in the form nolanlabs accepts.

<!doctype html>
<style>
  body { margin: 0; display: grid; place-items: center; height: 100vh; }
  h1 {
    color: #fff; font: 700 140px sans-serif;
    animation: rise 0.8s cubic-bezier(.2,.7,.2,1) both;
  }
  @keyframes rise { from { opacity: 0; transform: translateY(24px); } }
</style>
<h1>Salt</h1>

Take it at 24 fps:

  1. Frame 12 is 12 / 24 = 0.5 s. The renderer pauses the CSS animation and sets its time to 500 ms. It doesn’t wait half a second.
  2. The rise ends at frame 19.2, so frame 20 is the first frame fully at rest. Longer rise, fewer frames of hold: you can reason about this in frames, the way you do with a cut.
  3. The picture is a PNG with alpha. Nothing paints the background here, so the footage under the clip shows through.
  4. To change “Salt” to “Salt Flats”, you change one word and that clip redraws. The rise, type and position stay as they were.

The same approach scales up through GSAP, anime.js, Lottie and Three.js, where the renderer moves each library to the exact time instead.

How it differs from Remotion and HyperFrames

All three draw a browser page frame by frame. The difference is what the agent writes and where the result lives. Remotion and HyperFrames are frameworks that also ship their own editing surfaces. nolanlabs, in the last column, is ours: an editor where the clip sits on the timeline.

Remotion HyperFrames (HeyGen) nolanlabs (ours)
Agent writes React components in a project HTML, CSS and JavaScript files HTML, CSS and JavaScript in a folder in the project
How time is set Your code reads the frame number and returns values for it Paused animations are seeked to each frame The page’s clock is replaced, and its animations are seeked to each frame
What breaks it Animation not driven by the frame number: it renders across browser tabs that share no state, so it flickers Wall clocks, unseeded randomness, loading things mid-render A page that reads the real clock another way, uses the network or keeps state the clock doesn’t drive. A check reports it
You get A rendered file, plus Remotion Studio A rendered file, plus a browser editor and a Studio app A clip on the timeline: trim, fade, blend, key its params, undo
Transparency WebM or ProRes 4444 WebM, ProRes 4444 MOV or PNG sequence ProRes 4444, WebM or PNG frames

Licenses, as of November 2026: Remotion is free for individuals, teams of up to 3 and non-profits, with a paid company license above that. HyperFrames is Apache 2.0.

Which fits? If your team already writes React, Remotion gives you your own components, and it works with coding agents too. HeyGen’s own comparison, written by a competitor, says Remotion is older, has more templates and production history, and has a simpler model to hold in your head. HyperFrames suits you if you have web material already, or want one file without a build step. In fairness to the source: nolanlabs adopted its data-start and data-duration convention. Choose nolanlabs if you want the graphic and the cut in one place, from the same chat.

How it differs from generated video

A generated clip is pixels with no source behind them. That’s the whole difference, and it cuts both ways.

Page drawn by frame Generated video
Words Real type, spelled as written Letters drawn by the model; check them one by one
Fixing one word Edit the word, redraw Generate again, and the rest may change with it
Length Whatever the clip needs Veo 3.1: 4, 6 or 8 s. Runway Gen-4.5: 2 to 10 s
Repeating a result Same inputs, same frame Veo documents a seed that aims to reproduce a result when nothing else changes. Change the prompt and that no longer helps
Photographic shots No Yes, and it’s why you’d use one

Those specs are as of November 2026, and they move often. Where a model wins is obvious: a shot of something that was never filmed. A page can’t do that, and neither can nolanlabs, which doesn’t generate footage. A night you shot by day is another matter: that’s a grade. For the wider question, see prompt box or timeline.

Where this falls short

These are limits of drawing in a browser, and of our version of it, from our known limits.

  • It can’t sit inside a moving shot. There’s no camera tracking to fix a graphic to a shot that moves.
  • It can’t go behind a person on a busy background. Mattes for motion clips come only from keying a green screen on an input clip, or from an input picture’s own alpha.
  • Heavy scenes are slow. With no graphics card, a Three.js world at 1080p takes one to three seconds a frame. Frames are kept, so it happens once per change. 4K is four times the work.
  • It leaves some handoffs behind. FCPXML, EDL and AAF drop motion clips. Export the clip with alpha instead, which is its own topic.
  • Words on the page aren’t titles. The editor’s title checks don’t see them. A separate motion check looks for words outside title-safe, cut-off words, missing faces and low contrast.
  • A keyed param moves only if the page reads it each frame. Most starters read theirs once, so keying them does nothing.
  • Taste is still yours. Say what the typeface, pace and feeling should be, in the chat or in your style notes.

How it works in nolanlabs

Ask in the chat. The agent starts from a starter (lower third, kinetic title, end card, chart build, and others) or writes the page whole, and the clip lands on the timeline as part of its proposal. A plain title doesn’t need any of this: the editor has native titles with their own looks, so you can ask for one of those.

Editing a composition writes a new revision rather than changing the old folder, so undo and a discarded proposal bring back what was there. Nothing is fetched while drawing, and each composition holds its own copies of its libraries and fonts. The title of Salt, our sample film, came from one chat message. See motion clips in the docs.

Blog

Keep reading.

Every piece

Written by the people who make nolanlabs, a video editor that edits with you.

Download for Mac Download for Windows Download

Free to start, on Mac and Windows.