The problem here is that Glowforge is using a fairly outdated version of AI image generation models. This is something that’s been solved in newer generations of the tech, but Glowforge hasn’t admitted exactly which engine they’re using so we can’t even evaluate if an upgrade is on the horizon or not. This isn’t like the rest of glowforge’s software, where it can stay in one place and not feel completely out of date for a long while. These models are changing so quickly that you’ll get left behind within three to six months if you try to bring an image generator to market like this.
The smart money is that they chose a system that they can swap out when newer generations come along, but newer generations are already here (and so much better), so I’m not sure what the holdup is.
The specific issue in this case is that the these large image models generally crop all of the images they absorb into squares before processing them, which leads the model to believe that a lot of pictures of people have the heads clipped off at the top.
In the meantime, there are some tricks. Try using “wide angle” or “portrait", sometimes that will skew the model toward looking at images that were not cropped when evaluating. You might also try “from Instagram”, believe it not. Instagram has required a square image from the beginning so the images it pulled from there shouldn’t be cropped.
It’s a long shot but maybe worth giving it a chance.
The latest iteration of stable diffusion, 2.0, has largely fixed the cropped portrait issue, as has MidJourney, and probably some of the other more active players in this space. Pure speculation, but magic canvas feels like it’s based on stable diffusion 1.5, just judging by its overall look and feel. We’ll see if GF tries to keep current.