Assuming all pictures are same orientation / angle - maybe use some facial dimensional metrics to decide by finding an expected mean for each distribution (cute

Assuming all pictures are same orientation / angle - maybe use some facial dimensional metrics to decide by finding an expected mean for each distribution (cute / not cute) during round one. Then analyze each pets distance from the expected mean for each metric and combine the expected distances. Use the combined deviation from the expected value for cuteness to rank all the pets.. You could use confidence intervals to require further rounds if the metrics aren't conclusive.