Skip to content

How the Feega canvas got fast and always sharp

This is how the Feega canvas went from stuttering to holding the display cap, and from blurry mid-zoom frames to sharp ones. Every number below was measured on one real user canvas: 53 pages with components. Desktop is an M2 Pro with a 60 Hz display. “Phone” is the same machine with a 4× CPU throttle.

The renderer is a Rust core built on vello, running on WebGPU and compiled to wasm.

To keep pans and zooms cheap, it keeps a pyramid of 256 px tiles. While the camera moves, it shows cached rasters instead of redrawing the scene. Tiles are cheap when the camera moves: moving a tile costs almost nothing.

Their problem is magnification. A tile rendered at one zoom and shown at a higher one is blurry until it is baked again.

To have tiles ready before they are needed, the core predicts where the zoom is going. It extrapolated zoom as (now / was)^6.

A fast zoom-out made that ratio small, and the sixth power made it close to zero. A near-zero zoom covers the whole world, so the prediction enumerated hundreds of millions of tiles. The tab froze.

The fix has three parts:

  • clip the prediction to the canvas bounds;
  • bound it to 2× the current zoom;
  • cap the tiles enumerated per frame.

Selection handles and labels are drawn by the core. They were rebuilt and resent on every zoom step: JSON.stringify in TypeScript, a parse in wasm, then a re-encode.

During a gesture they do not change, so now they are held still until it ends.

MeasureBeforeAfter
Desktop zoom fps (60 Hz cap)4860
Frames over 8.3 ms (uncapped)5525
p95 frame time (uncapped)18.6 ms4.4 ms
4× CPU fps3859.5

Tile bakes are now sized to a 4 ms budget per frame, so baking never eats the frame. At 4× CPU uncapped, fps went from 122–136 to 143–147.

The budget did not cover everything a gesture frame produces. The never-blurry fallback, which draws a region at full resolution when no sharp tile covers it, drew whole screen bands in one frame: 50–330 ms each on the real-sites canvas while zooming. Bakes, the fallback and the look-ahead now share one deadline per frame. During a gesture, tiles are baked one at a time and the fallback is drawn in 256 px blocks that carry on the next frame. Until a region is ready, the coarser level underneath is shown, stretched. Once the gesture stops, the canvas still sharpens fully.

Zoom across octaves, real-sites canvas, 120 HzBeforeAfter
Presented fps88.8–93115.4–119.9
Frames over 16.7 / 33 ms119 / 567 / 2
Longest presented gap342 ms108 ms

The core has a fast patch path that redraws only the edited part. It declined every edit inside a component instance, so on a canvas built from components, every edit rebuilt the whole frame.

After the fix, all edits patch: 10 out of 10 in the test set.

Edit (desktop p50)BeforeAfter
Nudge119 ms12 ms
Colour101 ms10 ms
Hide70 ms13 ms

Separately, by-id lookups inside loops were replaced with Map and Set indexes. One example: the layers panel on 2,000 layers went from 118 ms to 9 ms.

Tiles were fast but could never be sharp mid-zoom. So we built what Figma does, as an opt-in path: draw pages from their vector encodings at the current zoom, every frame.

  • The scene is built once per gesture, then redrawn each frame with the camera transform.
  • Effects (blur, shadows) are expensive to redraw. They are cached per power-of-two zoom level and rendered finer than needed, so they are only ever downscaled, never magnified.

Results on desktop:

  • 60 fps, the display cap;
  • no blurry frames;
  • mid-zoom sharpness up: Laplacian-variance median 234 → 260, p90 354 → 396;
  • 0 page dropouts in 9 runs.

The honest limit: full scene rebuilds still spike. At 120 Hz the rebuild frames cost 114–173 ms, and zoom on a real-sites canvas fell from 118–120 to 90–95 fps. So tiles are the default again on every device, and direct is opt-in with ?directZoom=1 until per-page rebuilds land.

  • Measure on the real canvas. The synthetic fixture hid both the component bug and the freeze.
  • Profile per phase before optimising. Three PRs of layout work gained nothing. Phase marks then showed the real cost was an O(n²) scan.
  • Check what an fps number measures. The earlier 48 fps cap was screen recording: the harness, not the product.