[✨ feature request] Is "change of emitter during rendering"available?

[zhaoguangyuan123]

Summary

Just want this simple feature.

init_images = [mi.render(scene, emitter=emitters[i], spp=128) for i in range(emitter_count)]

From the document and tutorials I can only see that the "sensor" but not the "emitter" is able to change.
Is such feature already available or any other workarounds??

Looking forward to hearing any suggestions from you.

Best
Guangyuan

[merlinND]

Hello @zhaoguangyuan123,

Such an argument in the mi.render() interface wouldn't make a lot of sense, because a scene can have an arbitrary amount of emitters.
Instead, you could e.g. create a scene with all emitters in it and use the mi.traverse() mechanism to update the value of the emitters before rendering. Something along the lines of:

params = mi.traverse(scene)

def enable_only_emitter(name):
    for k, _ in params.items():
        if k == name + '.radiance.value':  # Check exactly how the relevant property is named in your scene
            params[k] = 1.5  # Will also depend on your scene
        else:
            params[k] = 0.
    params.update()

for i in range(emiter_count):
    enable_only_emitter('emitter_' + i)
    img = mi.render(scene, spp=128)
    # ...

Of course you'll have to adapt the naming convention and the parameter data types according to what's in your scene.

Another option is to create the scene within each loop iteration using load_dict and giving only the emitter you want in that iteration.

[merlinND]

I am not sure I understand the purpose of having ImageNet images as emitters in a 3D scene.
What are the other objects in the scene, should they be illuminated by that emitter?
If there is no other object, then is rendering even needed? (could work with the image itself directly in 2D instead)

Anyway, yes it would be achievable by using a textured AreaLight emitter.
At each iteration, you could use params to set the values of the emissive texture to the new image from ImageNet.

[zhaoguangyuan123]

I am not sure I understand the purpose of having ImageNet images as emitters in a 3D scene. What are the other objects in the scene, should they be illuminated by that emitter? If there is no other object, then is rendering even needed? (could work with the image itself directly in 2D instead)

Anyway, yes it would be achievable by using a textured AreaLight emitter. At each iteration, you could use params to set the values of the emissive texture to the new image from ImageNet.

Sure. Thank you.

You are right that ImageNet is only 2D. Ideally, I want to load 3D dataset but my thought was to start with ImageNet, as the 2D object is relatively easier.

I just come up with another possible way. For the same task, can I just initialize a object in between the light source and sensor the while replace the object in the scene in every rendering and keep the emitter to be fixed?

[zhaoguangyuan123]

What are the other objects in the scene, should they be illuminated by that emitter?

Yes in my previous plan the other objects in the scene are the object to optimize for and they are also illuminated by the emitter.
Let me know if this does not resolve your confusion.

[merlinND]

replace the object in the scene in every rendering

Unfortunately, that's actually not supported at the moment, we cannot insert or remove shapes once a scene has been loaded.

the other objects in the scene are the object to optimize for and they are also illuminated by the emitter.

Okay, in that case I think that the textured area emitter would work!

[zhaoguangyuan123]

Sure. I will go ahead and try the "Arealight" plan you suggest.
Thank you, Merlin!