Separate output for each wavelength

[UNakade]

Is that possible in Mitsuba 3?

I tried to use specfilm with very narrow channels, but the output was bizarre. Could someone help me with an explanation?

I described a scene in which the camera is looking directly at an area emitter which has a constant spectrum (defined by [(360, 1), (595, 1), (830, 1)]). Then I defined a specfilm with 93 narrow channels, each of which is given by [(wavelength - 5, 0), (wavelength, 1), (wavelength + 5, 0)] (a narrow triangle shape). If I then used the path integrator, the output was 0. But with a volpath integrator, there was some output. I took an average of each channel and the results varied wildly (with the largest value almost three times the smallest). I repeated the same thing in a different Python shell, but the result was exactly the same. Could someone tell me what on Earth is going on?

Here's the code:

import mitsuba
from mitsuba import ScalarTransform4f as Transform
import numpy as np
import matplotlib.pyplot as plt

film = {
    'type': 'specfilm',
    'width': 512,
    'height': 512
}

for wavelength in range(365, 830, 5):
    film[f'{wavelength}'] = {'type': 'spectrum', 'value': [(wavelength - 5, 0), (wavelength, 1), (wavelength + 5, 0)]}


scene = {
    'type': 'scene',
    'integrator': {
        'type': 'volpath',
        'max_depth': -1,
        'rr_depth': 20
    },
    'emitter_area': {
        'type': 'rectangle',
        'bsdf': {
            'type': 'null'
        },
        'emitter': {
            'type': 'area',
            'radiance': {
                'type': 'spectrum',
                'value': [(360, 1), (595, 1), (830, 1)]
            }
        }
    },
    'camera': {
        'type': 'perspective',
        'to_world': Transform.look_at(
            origin=[0, 0, 1],
            target=[0, 0, 0],
            up=[0, 1, 0]
        ),
        'fov': 150,
        'sampler': {
            'type': 'independent',
            'sample_count': 150
        },
        'film': film
    }
}


mitsuba_scene = mitsuba.load_dict(scene)
output_image = mitsuba.render(mitsuba_scene, spp=256)

output_image_numpy = output_image.numpy()
mean_output_image_numpy= np.mean(output_image_numpy, axis=(0, 1))

plt.plot(range(365, 830, 5), mean_output_image_numpy)
plt.ylim([0, 1.1*np.max(mean_output_image_numpy)])
plt.xlabel('wavelength')
plt.ylabel('average of channel')
plt.show()

And here's the plot:

The channels are supposed to be like this:

[UNakade]

This jagged output seems to be fixed in commit 39ff404. Its parent commit (d456826) seems to have the issue (I checked).