AI image generation just got faster and more diverse — without needing to predict rewards ahead of time

Researchers replaced a standard resampling technique in AI image generators with a method inspired by population biology, which prevents the models from repeatedly copying the same successful outputs. In practice, this means image generation can run faster while producing more variety, without requiring the model to pre-calculate how good each image will be.

Image generation models have a recurring problem: when you try to steer them toward better outputs, they collapse into repetition — the model keeps copying the same winning path instead of exploring alternatives. This paper fixes that collapse by using a different mechanism for which candidate images survive to the next step. The speedup matters because inference compute is increasingly the bottleneck in deployed AI systems; cutting generation time in half while improving quality is a meaningful shift in the cost-to-performance curve for commercial image generation.