I don't know, whatever works best and smoothly with Constructs behaviors. The current problem is that for high res art that isn't blocky, you'll need to come up with lots and lots of collision tiles that you'll have to paint manually, cause the sprites can be a bit more complicated - which takes up a LOT of time.
So the idea is that Construct would analyze the sprite shape and generate a simplified version by itself.
The current per pixel implementation is a bit 'too perfect' - which means that it detects every little variation in the sprite - it there's a 1 pixel dot standing out, the playerSprite with platform behavior will often come to a complete halt when it collides with the pixel. So you have to do it manually and make sure that everything is perfect.
So the optimal implementation would be like a channel that you could turn on in the sprite editor and then change the collision mesh of a sprite with a couple of sliders - like changing how much it simplifies the shape, offset it so that it looks like the player sprite would slightly stand inside of the sprite, etc. - just a procedural way to create a collision mesh / sprite for our tiles.