I'm working on a diffusion inference engine. I dislike node editors like ComfyUI and wanted a decent python API to control it. The first version was another python/PyTorch variant which was decent but still way too slow to my liking.
I ended up rewriting all the inference in C++/CUDA and I'm finally getting close to limits on the devices I tested. The benchmarks are still a work in progress, I plan to add high resolution comparisons (1536^2 for images, and 1344x768 for video) to give insight in how well this scales.