Diffusion models generate stunning images, but they're slow. A single image requires dozens of sequential neural network evaluations — each one a full forward pass through a U-Net. DPM-Solver++ brought that down to 10-20 steps with reasonable quality, and it's the current state of the art. But what if we could do better by borrowing techniques that the scientific computing community has used for decades?