You can get around many these compatibility issues by creating a custom element that inherits from HTMLImageElement. This provides API compatibility. For CSS compatibility, the elements you would replace in a MutationObserver would be the same tag name but a different namespace for CSS compatibility.
For the CSS compatibility trick, see https://eligrey.com/demos/hotlink.js/ which replaces images with CSS-compatible (not HTMLImageElement-compatible) iframes.
> - It is very hard to check if a certain image is visible or not, and react accordingly.
You can use Element.checkVisibility()¹ and the contentvisibilityautostatechanged event²𝄒³ to do this. Browser support is currently limited to Chromium-based browsers.
1. https://drafts.csswg.org/cssom-view/#dom-element-checkvisibi...
2. https://github.com/vmpstr/web-proposals/blob/main/explainers...
3. https://caniuse.com/mdn-api_element_contentvisibilityautosta...
Canvases are rectangles. The viewport is a rectangle. Checking if rectangles overlap is easy.
- Rendering images from a different origin without CORS. This is a fundamental limitation of any JS or WebAssembly solution and can't be fixed. Thankfully this use case is relatively rare.
- Not all approaches can provide a seamless upgrade. For example if you replace all `<img src="foo.jxl">` with a canvas DOM will be changed and anything expecting the element to be HTMLImageElement will break. Likewise, CSS Painting API [1] (a relevant part of CSS Houdini) requires you to explicitly write `paint(foo)` everywhere. The only seamless solution will be therefore a service worker, but it can't introduce any new image format; it can only convert to natively supported formats. And browsers currently don't have a "raw" image format for this purpose. JXL.js [2] for example had to use JPEG as a delivery format because other formats were too slow, as I've been told.
- It is very hard to check if a certain image is visible or not, and react accordingly. This is what I intended to imply by saying that canvas has to unconditionally retain all pixels, because if implementations can't decide if it's safe to unload images, they can't do so and memory will contain invisible images in the form of canvases. Browsers do have a ground truth and so can safely unload currently invisible images from memory when the memory pressure is high.
[1] https://developer.mozilla.org/en-US/docs/Web/API/CSS_Paintin...
[2] https://github.com/niutech/jxl.js