We compare our gradient path tracing (G-PT) method using L1 and L2 reconstruction to ordinary path tracing (PT). We show renderings with twice as many base samples for PT compared to G-PT. This provides an approximately equal time comparison, since the overhead of G-PT per base path is about a factor of 2.5. In other words, at exact equal render time the PT images would be slightly less noisy. The reference images were rendered with bidirectional path tracing. We measure the error as relative mean square error (relMSE).

Instructions: Click on the thumbnails to show the images at their original resolution.

Bathroom 2048spp

	G-PT L1 relMSE: 0.001998 PT (4096spp) relMSE: 0.012233	G-PT L2 relMSE: 0.002095 Reference 40000spp

Bathroom 256spp

	G-PT L1 relMSE: 0.009968 PT (512spp) relMSE: 0.091963	G-PT L2 relMSE: 0.011023 Reference 40000spp

Bookshelf 2048spp

	G-PT L1 relMSE: 0.000763 PT (4096spp) relMSE: 0.002943	G-PT L2 relMSE: 0.000887 Reference 25000spp

Bookshelf 256spp

	G-PT L1 relMSE: 0.003247 PT (512spp) relMSE: 0.020531	G-PT L2 relMSE: 0.003965 Reference 18000spp

Kitchen 2048spp

	G-PT L1 relMSE: 0.001455 PT (4096spp) relMSE: 0.006881	G-PT L2 relMSE: 0.001701 Reference 32000spp

Kitchen 256spp

	G-PT L1, 256spp relMSE: 0.007584 PT, 512spp relMSE: 0.051185	G-PT L2, 256spp relMSE: 0.009138 Reference 32000spp

Sponza 64spp

	G-PT L1 relMSE: 0.000804 PT (128spp) relMSE: 0.011792	G-PT L2 relMSE: 0.000888 Reference 24000spp

Sponza 8spp

	G-PT L1 relMSE: 0.006616 PT (16spp) relMSE: 0.094282	G-PT L2 relMSE: 0.006455 Reference 24000spp

Veach-Door 2048spp

	G-PT L1 relMSE: 0.004395 PT (4096spp) relMSE: 0.029148	G-PT L2 relMSE: 0.008662 Reference 32000spp

Veach-Door 256spp

	G-PT L1 relMSE: 0.021791 PT (512spp) relMSE: 0.223310	G-PT L2 relMSE: 0.051629 Reference 32000spp

This webpage was built using the code found on: http://www.cs.huji.ac.il/~raananf/projects/lss_upscale/sup_images/index.html.