the spir-v here is automatically generated from the oidn onnx.
this module requires a .dat weights file in ~/.config/vkdt/data containing
the pre-trained weights for the oidn ldr network variant.
these can be downloaded from here.
if the input exposes some kind of inter-pixel correlation (such as from a
preceeding demosaicing operation) this module will not do anything at all. oidn
will classify the correlated noise statistics as signal and pass them on
unchagned. if you have a very noisy mosaic image, use half size as
demosaicing method to keep pixels uncorrelated. avoid resampling nodes
before denoising.
arch select the gpu achitecture. auto means use the vendor of the current gpu. others may or may not work on your device.model select the oidn network modelinput scene-referred data, preferrably low-res and in [0,1]output the denoised imageit is my understanding that oidn has been trained on purely synthetic data, output from rendering systems. thus, no internet scraping or violation of rights of third parties took place during training.
the spir-v found in this module is generated from onnx using the denox compiler, in particular the denox vkdt codegen. the method is described in the paper: Karl Sassie, Johannes Hanika, Lucas Alber, Reiner Dolp, and Carsten Dachsbacher. Optimizing vulkan dispatch schedules for real-time U-net denoising. Proc. ACM Comput. Graph. Interact. Tech., 2026. pdf