Based on a conversation on Discord.
- using a perceptive image hash
- baking to a UDIM/giant map and then using perceptive hashes to see how close the results for one rectangle are to another
- merge rectangles which are “too similar” so they use up less space on the map (potentially freeing up space for something else.) Although perception will change based on size of rectangle: high resolution will not look like small resolution but the smaller resolution of each box the more they will start looking the same.