Bad AI denoiser for caustics at PT on OR 2022.1

[nuno1980]

Hello.

See 3 images:
-PT (path tracing) kernel at 4M samples per pixel (= 4 laps of regions picked bottom of 1MS/px)

-PT with denoiser (same MS/px) - look highlights

-new Photon Tracing kernel with denoiser at 515 S/px but no any variable caustics (less than 0.3 roughness ground)

We don't want any blur nor any invisible for caustics for PT kernel but AI denoiser improved for caustics in next version of OR.

[elsksa]

Hi,

The rule of thumb with (nearly all?) denoisers is to provide them an as cleaned input as possible (some yield superior results with additionally provided AOVs).

I'd be curious to see the Photon Tracing Kernel settings. On such a simple scene, it shouldn't be struggling to the point of requiring the use of the denoiser. It may, for a "polishing pass" rather than a means to an end, especially on such hardware.

How come the Photon Tracing Kernel has literally identical stats as the two other PT renders/screenshots?

[nuno1980]

Hi,

The rule of thumb with (nearly all?) denoisers is to provide them an as cleaned input as possible (some yield superior results with additionally provided AOVs).

I'd be curious to see the Photon Tracing Kernel settings. On such a simple scene, it shouldn't be struggling to the point of requiring the use of the denoiser. It may, for a "polishing pass" rather than a means to an end, especially on such hardware.

How come the Photon Tracing Kernel has literally identical stats as the two other PT renders/screenshots?

I deceive you - I'm very sorry for the lack of the image rendered at Photon Tracing. lol

PT settings:

Photon Tracing settings - see OP