Why does rendering produce negative values?

[Disgruntoad]

According to the Emitters page (https://mitsuba.readthedocs.io/en/latest/src/generated/plugins_emitters.html):

Whether Mitsuba is rendering in spectral or RGB mode, if the final output is an sRGB image, it can happen that it contains negative pixel values or be over-saturated. These results are left un-clamped to let the user post-process the image to their liking, without losing information.

According to the HDRFilms section from the Films page (https://mitsuba.readthedocs.io/en/latest/src/generated/plugins_films.html):

It stores the captured image as a high dynamic range OpenEXR file and tries to preserve the rendering as much as possible by not performing any kind of post processing, such as gamma correction—the output file will record linear radiance values.

I've noticed in some of my own renders written as rgb to exr that the pixel value may be negative. I checked the scene bsdf file and none of the spectra were negative.

How should I interpret / understand what negative values mean?

[wjakob]

Negative sRGB values are outside of the gamut of representable colors in sRGB. This is an issue that is unique to RGB colors and will not show up with spectral representations.

[Disgruntoad]

Ahh, good to know. I assume that if we want to avoid this issue in the hdrfilm plugin, the XYZ format is used to provide a standardized representation covering the full colour space?

Follow-up question -- what is the purpose of rendering the film with an alpha channel? Does this have any behavior in mitsuba or is it just to provide a common format?

[njroussel]

I assume that if we want to avoid this issue in the hdrfilm plugin, the XYZ format is used to provide a standardized representation covering the full colour space?

Yes, XYZ would guarantee that everything can be mapped.

what is the purpose of rendering the film with an alpha channel? Does this have any behavior in mitsuba or is it just to provide a common format?

Integrators' sample method returs both the incident radiance for a light path and a boolean. The latter is typically used to indicate if the ray had any interaction, i.e: did it hit a surface. The accumulation of these boolean will constitute your alpha channel. For example, if you render the mi.cornell_box() scene, the sensor's field of view is slightly larger than the box which is why the final rendered image has a small ~3 pixel transparent border.