I recently investigated the performance of the AI denoiser filter and I was wondering if the docs should have some reference numbers? This would allow the users to more easily understand if their integration is hitting some slow path or is the performance expected.
In my case I compares RIF AI denoising filter with NVIDIA OptiX one. On the 3080 Ti graphics card I see that processing an image of the size 1000x600 takes around 11ms. For comparison OptiX takes at most 500 us.
This is a bit puzzling since the architecture I see when loading the "denoise_c9_ldr.pb" model into tensorboard seem to be similar if not simpler than the one described for optix in this paper.
The performance matches the sample app.
I also feel like it's worth mentioning weather this denoiser is suitable for real-time applications.
I recently investigated the performance of the AI denoiser filter and I was wondering if the docs should have some reference numbers? This would allow the users to more easily understand if their integration is hitting some slow path or is the performance expected.
In my case I compares RIF AI denoising filter with NVIDIA OptiX one. On the 3080 Ti graphics card I see that processing an image of the size 1000x600 takes around 11ms. For comparison OptiX takes at most 500 us.
This is a bit puzzling since the architecture I see when loading the "denoise_c9_ldr.pb" model into tensorboard seem to be similar if not simpler than the one described for optix in this paper.
The performance matches the sample app.
I also feel like it's worth mentioning weather this denoiser is suitable for real-time applications.