Upscaled models using the Block Expansion method. Unlike the more common DUP Scaling, BE doesn't require fine-tuning to recover lost performance.