The object returned by the function expresses a tiny subset of the properties of a DOM node. Often just {className, childNodes: [...]}. Only those explicit attributes are checked for update or otherwise dealt with by my code. My code has no idea that a DOM node has a thousand properties. By contrast, a ReactComponent is more complex from a JS POV than a native DOM node.

In other words, if my code returns: {className: 'foo'} at time t0, and then returns {} at time t1, the className in the DOM will be 'foo' at t0 and at t1. That is not at all how exhaustive tree diffs work, and not at all how react works.

With 5,000 nodes, you might have 8K-15K property comparisons. Per-render CPU load thus grows linearly and slowly with each new node. I can re-render a thousand nodes in 2-5 milliseconds with no framework churn or build steps or any of that. But more importantly, we have the ability to step into "straight-to-canvas" mode (or whatever else) without rupturing any abstractions and without awkward hacks.

This is unidirectional data flow plus component-based design/organization while letting the DOM handle the DOM: no fake elements, no fake events -- nothing but utterly fast strict primitive value comparisons on shallow object properties.

EDIT: Earlier I said that a node changes if and only if its state description changed; that is not strictly true. "if and only if" should just be "only if".