At the same time, I realised the Human Pnagenome Reference Consortium had made a second release of genomes.
So I rebuilt release 1 without masking, and added a release 2 database with no masking. The improvement in detection accuracy was substantial:

However, the computational resource usage (runtime/memory) of LRGE was MUCH better than assembling

We benchmarked >3,000 bacterial genomes and found that LRGE (our method) achieves significantly better accuracy than k-mer-based methods like Mash and GenomeScope and performs on par with full genome assembly (Raven)

6/ And it’s efficient! ⚡
LRGE uses significantly less CPU and memory than traditional approaches, making it ideal for both high-performance clusters and resource-limited environments.

5/ LRGE vs. the competition 🔥
LRGE delivers estimates as reliable as assembly-based methods and better than k-mer-based approaches.
Relative error (y-axis) measures the proportional difference between the estimated and true genome size.