Infinite canvas

An infinite canvas has no page and therefore no end. Content sits at coordinates on an unbounded plane, and what the reader manipulates is not the content’s position in a document but the camera: a bounded viewport that pans and zooms over the plane. Whiteboards, node editors, design tools and diagramming apps all converge on this model for the same reason, which is that the work being done is spatial. Where two things sit relative to each other is the meaning, so a layout that reflows would destroy the document.

The difference from a scroll container is worth stating precisely, because they look alike from a distance. A scroller has a content extent: it knows how tall its content is, the scrollbar is a proportion of that height, and scrolling to the bottom is a real place. A canvas has neither. Its content has a bounding box, but the plane it sits on does not end, so there is nothing for a scrollbar to be a proportion of. Everything that follows comes from that one fact. There is no home to return to unless the application invents one, which is what a “fit” or “zoom to fit” command is: it frames the bounding box of what exists, which is the only privileged view a canvas has. And a reader can get genuinely lost, in a way a scroller makes impossible, which is why a minimap is close to mandatory rather than a nicety. The overview badge in the corner of every canvas tool is one.

The gestures are the other half of the term. Pan moves the camera and pinch to zoom changes its scale, and the hard part is that the same gestures also mean “move this object” and “select these objects”, so a canvas spends most of its interaction budget on disambiguating a drag: space to pan, a hand tool, middle drag, two fingers, or a drag that starts on empty plane rather than on a shape. Zoom must keep the point under the pointer fixed, since a zoom that recentres throws away the reader’s place. Below some scale the shapes stop being readable and a good canvas switches to drawing what the shapes mean instead, the same trick a minimap plays.

Two costs are easy to underestimate. Rendering is one: an unbounded plane means culling what is off camera, in the same spirit as list virtualization but in two dimensions. Accessibility is the other, and it is the harder one. A plane of freely placed shapes has no reading order, so keyboard traversal, focus order, and anything a screen reader announces have to be invented from the geometry or from a structure kept alongside it. The tools that do this well tend to maintain an explicit list view of the canvas rather than deriving one on the fly. tldraw publishes its camera and shape model as a library, which is a useful thing to read before building a third one.