Oh I already answered you on Lee's site but here it is copypasted again if anyone else is interested:
It depends on the exact implementation but this is basically what I would do.
1. Generate a single atlas from the thumbnails.
2. Allocate an empty texture large enough to hold any one of the zoomed images.
3. If you click a thumbail you don’t need to dispose/create anything just upload the zoomed image to the free allocated texture already on GPU.
This way you can render it with 2 draw calls all the time, no allocation or disposing. You may end up with more atlases if there are a lot of thumbnails, or if you have paging dynamically reupload new thumbnails to the atlas etc. But you get the idea.