We've put out a paper on scheduling optimizations for our ONNX2Vulkan generator for U-Nets.
For OpenImageDenoise, the generated Vulkan implementation is competitive with established inference frameworks and OIDN's own CUDA backend, while integrating directly into Vulkan renderers.
lalber.org
Optimizing Vulkan Dispatch Schedules for Real-Time U-Net Denoising
We present a light-weight and portable ONNX to Vulkan conversion framework that is designed for efficient, fully GPU-resident deployment of U-Net-based networks.
