Detailed Comparison of I1Profiler, Argyll and Profile Maker 5

[Doug Gray]

These charts from comparing profiles made with 6 different patch counts using the 3 profile creation tools on each set. 4 printer/paper combinations were used. Accuracy was determined for 848, random patches distributed over the printer/paper gamut and 192, near neutral patches randomly chosen near the gray axis.

The printer/paper  combinations were: Costco Glossy for the Epson 9800, Canon Pro1000, and Canon 9500 II. Matte (MP101) with the Pro1000 was also tested. All were done at the highest quality print settings. All profiles were also created using the highest quality settings.

Generally, the I1Profiler and Argyll were quite close with the Argyll give slightly better results with larger patch sets and I1Profiler doing the best at smaller sets. The differences were fairly minimal. Profile Maker 5, however is consistently inferior.

All of this data is from the 8 page, 957 patch each targets described in more detail in the crowdsourcing thread.


This is the performance across the full printer/paper gamut:



These are the results for the near neutrals:

[GWGill]

The fitting used by Argyll is tuned (i.e. the regularization or smoothing level) differently for different patch set sizes. The simulation/tuning runs I used were laborious, and difficult to pin down given the variation in random error simulation, and I think for smaller patch sets I erred on the side of more smoothing rather than less. Playing with the colprof -r parameter would likely tweak the standing of the fit one way or another.

[Doug Gray]

The fitting used by Argyll is tuned (i.e. the regularization or smoothing level) differently for different patch set sizes. The simulation/tuning runs I used were laborious, and difficult to pin down given the variation in random error simulation, and I think for smaller patch sets I erred on the side of more smoothing rather than less. Playing with the colprof -r parameter would likely tweak the standing of of the fit one way or another.

Graeme, I think your description matches with these results. Further, on some of the posts on my crowdsourcing thread, there are dE00 for AtoB for the patch sets used to create the profile. I1Profiler, for small patch sets, does close to an exact match and doesn't cut in smoothing until the count gets fairly large. This supports your conjecture. This works well for well behaved printers like the Pro1000 but not so well for the 9800. The 9800 AtoB and BtoA tables are not isomorphic. There are some widely spaced device RGB values that have the same Lab vals in AtoB.