The two-stage strategy of first generating at low resolution and then performing a 2x upscale can circumvent the facial breakdown issues commonly seen with direct high-resolution generation.