Remove lighting importance sampling

[KovenYu]

Hi Team,

Thanks for sharing the wonderful work and amazing codebase! I wonder if it is possible to remove lighting importance sampling in the cuda kernel code, so that we only use BRDF importance sampling. My assumption is to remove:

// Light importance sampling sx = ((float)(params.perms[lightIdx][i] % params.n_samples_x) + uniform_pcg(rng_state)) * strata_frac; sy = ((float)(params.perms[lightIdx][i] / params.n_samples_x) + uniform_pcg(rng_state)) * strata_frac; ray_dir = lightSample(sx, sy, pdf_light); pdf_bsdf = bsdf_pdf(pDiffuse, pSpecular, gb_normal, wo, ray_dir, alpha); process_sample(idx, ray_origin, ray_dir, gb_pos, gb_normal, gb_view_pos, gb_kd, gb_ks, pdf_light + pdf_bsdf, sample_frac, diff, spec, diff_grad, spec_grad); diffAccum = diffAccum + diff; specAccum = specAccum + spec; From L512, and then remove pdf_light in L527.

Am I doing correctly to achieve my goal? I have no prior CUDA programming experiences. Appreciate very much for any comment!

Thanks, Koven

[JHnvidia]

Hi,

Spontaneously this seems correct. I think that the balance heuristic should simplify to normal sampling if you set pdf_light=0 (or remove it) in L527. It's notoriously easy to introduce errors in MC-rendering through, so it would be a good idea to forward render some reference mesh (e.g. using https://github.com/NVlabs/nvdiffrecmc/blob/main/dataset/dataset_mesh.py) at high sample counts with our original code and your modified version to verify that there are no intensity shifts. You can expect the noise to increase, but given high enough sample count the two versions should converge to the same image.

Light importance sampling is quite central to some of our results. Without it I would not expect scenes with strong directional lights or shadows to work as well.