Rendering engine · last reviewed 18 November 2026

Pavo AI Video: How the in-browser renderer works

No render farm, no model API, no queue. This page states exactly what happens between pressing Generate and downloading a WebM file — and where the limits of a local renderer are.

The pipeline, step by step

  1. Your still is loaded into an in-memory Image object. Uploaded files become a local blob URL and are never posted to a server.
  2. The prompt is scanned for motion words (push, pull, pan, drift, hold) and the studio suggests a matching camera path. You always have the final choice.
  3. For each frame, Canvas 2D draws the still with a time-dependent scale and offset — the six camera paths — then composites a slow light sweep, a vignette and an optional caption bar.
  4. canvas.captureStream(30) exposes those frames as a 30 fps MediaStream.
  5. MediaRecorder encodes the stream to WebM (VP9 where supported, otherwise VP8) while the clip plays in real time.
  6. When the last frame is drawn, the recorder stops, the chunks become a Blob, and an object URL powers the preview and the download link. Nothing is stored after you leave the page.

Because the still is transformed rather than redrawn, results are deterministic: the same image, prompt and settings produce the same motion every time, and faces, hands and geometry never morph.

What a release check covers

CheckPass condition
Frame geometryThe image covers the canvas at every t, in all three aspect ratios, with no exposed edges.
EndingThe final frame holds the composition long enough to be used as an edit point.
RecordingThe exported WebM plays in a fresh tab at the requested duration and frame rate.
PrivacyNo network request contains the uploaded image, prompt or generated clip.
HonestyNo page claims a diffusion model, quotes third-party prices or invents user reviews.
AccessibilityControls are labelled, the canvas has a text alternative, and status changes are announced.

The reviewed date appears on every page and in the sitemap. We re-check the pipeline whenever browser media APIs change.

What ships with the studio right now

12Shot presets
6Render fixes
5Help guides

The boundary, stated plainly

This renderer produces motion from a single still. It does not synthesise new objects, novel scenes or talking characters — that requires a diffusion-based video model. We say so on the Create page and in the FAQ, because a clear boundary is more useful than a vague claim. The failure-to-fix table covers the generative failure modes you may hit if you take a shot brief into such a model elsewhere.

The illustrations and sample stills on this site were generated for this project; they are not screenshots of any commercial product and not outputs of a third-party video model.

Corrections

If a technical statement here is wrong, misleading or out of date, use the contact page to tell us which page and which line. We correct factual errors rather than quietly rewriting them.

See it run

Press Generate once and the whole pipeline is visible.

The preview canvas shows every frame as it is drawn, and the resulting WebM is ready to download as soon as the clip finishes.