Specification of envmap.data without an image file.

[maxfrei750]

I'd like to optimize the data attribute of an envmap. To do so, it needs to be initialized (e.g. with zeros or ones). I can do this using an image file and the filename attribute. However, IMHO it would be more elegant to do the initialization in memory, e.g. creating a suitable tensor while loading the scene. I tried to plug the initial data into the scene dictionary using a nested list, a numpy array or a drjit tensor:

e.g.

{
    ...
    "environment": {
        "type": "envmap",
        "data": dr.ones(dr.llvm.TensorXf, (8, 8, 4))
    }
    ...
}

However, I always receive errors like

RuntimeError("\u200b[xml_v.cpp:337] Unkown value type: <class 'drjit.llvm.TensorXf'>")

Is there a way to do this?

Thanks!

EDIT: Related: Should the range of the data be [0, 1]?

[Speierers]

Hi @maxfrei750 ,

Currently it isn't possible to pass a TensorXf object as property. I will add the possibility to directly pass a mi.Bitmap object (which can be created from a TensorXf) to the `envmap plugin.

[Speierers]

I added this feature in this commit: 9389c8

[maxfrei750]

Thanks for the quick implementation. It works really nicely. Here's a small snippet for others to get started:

environment_size = 64

scene_dict = {
    "type": "scene",
    "environment": {
        "type":
            "envmap",
        "bitmap":
            mi.Bitmap(
                mi.TensorXf(mi.PCG32(size=environment_size**2).next_float32(),
                            shape=(environment_size, environment_size, 1))),
    }
}

It initializes the envmap with a random noise, which is preferable to a blank envmap, if I understood

mitsuba3/src/emitters/envmap.cpp

Lines 173 to 174 in ea513f7

	/* Be wary of constant environment maps: average and minimum
	should be sufficiently different */

correctly.