Differential Version of bitmap.convert(mi.Bitmap.PixelFormat.RGB, mi.Struct.Type.UInt8) for mi.TensorXf

[shallowtoil]

Hi,

I want to clip the rendered image from mi.render() to the range (0, 1) such that it can receive the gradient from off-the-shelf vision models with typical input range being (0, 1). But I noticed in mitsuba3 such transformation is only supported using mi.util.convert_to_bitmap() which is based on a non-differentiable Bitmap. I am wondering whether we have a counterpart directly for mi.TensorXf allowing me to do that.

I've noticed in the tutorial ** (1.0 / 2.2) is used to do that. Does this single approximated step serve the purpose?

Thanks

[njroussel]

Hi @shallowtoil

The Bitmap class, indeed, doesn't have any sort of gradient tracking capabilities. Also, I don't believe we provide any built-in functions to do the tonemapping/sRGB ramp differentiably.
The snippet you mention from the tutorial can technically serve that purpose. Again, a 2.2 gamma correction is just an approximation.