MIS Compensation: Optimizing Sampling Techniques in Multiple Importance Sampling

(supplemental material - rendered image interactive comparisons)

Here we show full image quality comparisons from our paper in an interactive way, where even subtler differences are clearly visible. You can move your mouse inside the big image to change what part of each image is shown. You can also rearrange images or change them by dragging a thumbnail from a list of available images above each comparison and dropping on the big image.

Compared variants

VariantUsed shortcut
ReferenceRef
Original MIS (multi-sample)MIS (multi)
Our normal-independent practical solution (multi-sample)Ours-ni (multi)
Our normal-dependent practical solution (multi-sample)Ours-nd (multi)
Normal-dependent sampling without the compensation (multi-sample)Normal (multi)
Resampled importance sampling (multi-sample)RIS (multi)
Original MIS (one-sample)MIS (one)
Our normal-independent practical solution (one-sample)Ours-ni (one)
Our normal-dependent practical solution (one-sample)Ours-nd (one)
Normal-dependent sampling without the compensation (one-sample)Normal (one)
Resampled importance sampling (one-sample)RIS (one)

Application I: Image-based lighting

Used HDR maps and corresponding pdfs

Low-constrast map

Used in the Synthetic and Car scenes.

Original map MIS pdf Our pdf

Mid-constrast map

Used in the Synthetic scene.

Original map MIS pdf Our pdf

High-constrast map

Used in the Synthetic, Pills, and Room scenes.

Original map MIS pdf Our pdf

Synthetic scene

Equal-sample comparison (1 sample per pixel)

MIS (multi) Ours-ni (multi)

Pills scene

Equal-time comparison (5 seconds)

MIS (multi) Ours-ni (multi) Normal (multi) Ours-nd (multi) RIS (multi) MIS (one) Ours-ni (one) Ours-nd (one) Normal (one) RIS (one) Ref

Equal-sample comparison (160 samples per pixel)

Note that the apparent brightness change inside the box between the reference and either compared method is caused by the originally singular fireflies eventually coming together and forming continous reflective caustics.

MIS (multi) Ours-ni (multi) Normal (multi) Ours-nd (multi) RIS (multi) MIS (one) Ours-ni (one) Ours-nd (one) Normal (one) RIS (one) Ref

Room scene

Equal-time comparison (50 seconds)

MIS (multi) Ours-ni (multi) Normal (multi) Ours-nd (multi) RIS (multi) MIS (one) Ours-ni (one) Ours-nd (one) Normal (one) RIS (one) Ref

Equal-sample comparison (160 samples per pixel)

MIS (multi) Ours-ni (multi) Normal (multi) Ours-nd (multi) RIS (multi) MIS (one) Ours-ni (one) Ours-nd (one) Normal (one) RIS (one) Ref

Car scene

Equal-time comparison (5 seconds)

MIS (multi) Ours-ni (multi) Normal (multi) Ours-nd (multi) RIS (multi) MIS (one) Ours-ni (one) Ours-nd (one) Normal (one) RIS (one) Ref

Equal-sample comparison (160 samples per pixel)

MIS (multi) Ours-ni (multi) Normal (multi) Ours-nd (multi) RIS (multi) MIS (one) Ours-ni (one) Ours-nd (one) Normal (one) RIS (one) Ref

Application II: Path guiding

Kitchen scene

Equal-time comparison (150 seconds)

MIS (one) Ours-ni (one) Ref

Equal-sample comparison (140 samples per pixel)

MIS (one) Ours-ni (one) Ref

Pool scene

Equal-time comparison (150 seconds)

MIS (one) Ours-ni (one) Ref

Equal-sample comparison (260 samples per pixel)

MIS (one) Ours-ni (one) Ref