I've compared the most recent versions of dispcalGUI (DCG)/ArgyllCMS, LightSpace, and CalMAN using the eeColor on a plasma display. The methodology was much the same as the OP comparison.
DCG/ArgyllCMS: 18.104.22.168 Beta/1.7.0_beta6
Some notable changes:
1. CalMan now uses single pass profiling to calculate LUT corrections, it retains the legacy multi-pass optimization method but that was not retested here.
2. I used DCG automatic profiling/LUT creation mode rather than a command line batch file to run the ArgyllCMS tools.
All three programs were driven with ~2600 profiling patches and took approximately the same amount of time to complete the process, which was between 1 and 1.5 hours using an i1display pro. All results except "LS-algo" represent standalone tests where profiling and verification measurements are performed within the same program.
Summary of verification statistics
LS-grid: This is a preset grid density of 14x14x14 (2774 points)
LS-opt: This is using the same optimized test set that was generated during the DCG auto-profiling mode (2527 points)
LS-algo: This is the display response using the LS algorithm with the DCG/Argyll measured profile.
Efficiency: Patch weighted ratio of native/post-LUT dE00 x 1000
The verification set consisted of 50 grayscale patches, 20 patch RGB luminance ramps, and 632 randomized patches. All values are dE00.
Case 1: Relatively good pre-lut response, average/worst 10% dE00: 1.65/4.0
All three programs performed well in Case 1 driving better than 90% of the verification patches below dE00 of 1. If I use the LS algorithm on the DCG measurement set the gap between LS and DCG narrows (Efficiency increase from 1.2 to 1.48 vs. 1.79). The only difference between the two data sets is a panel drift correction.
Case 2: Similar to pre-lut response of OP, average/worst 10% dE00: 3.92/7.75
The spread between the 3 programs increases as the pre-LUT response degrades, most significantly in the case of CalMAN. This can be further examined by looking at the LUT noise characteristics.
These are surface plots of the output-input code deltas for the red channel input video level 60 as a function of B and G input values.
Summary of LUT noise characteristics
Where RSS is the root-sum-square of the individual channel 1-sigma standard deviations.
In Case 1 the relative noise factor between CM and DCG is 3.1 and does not significantly impact the error distribution, but in Case 2 it jumps to 6.2 and degrades performance. [LS maintains a factor of 2 more LSB noise relative to DCG in both cases]
A further illustration is shown here which breaks out the 1-sigma noise level of the red channel as a function of blue input level.
It's not clear to me whether this is a result of issues with interpolation or with the LUT optimization itself.
Another interesting thing that can be seen in the surface plots is that the LS grid-based profiles are smoother than the ones using Argyll-optimized patches, and that the overall efficiencies between the two are not that different. This was not the case in the OP where for example the efficiency factor for that grid set was only 1.09 and it's now 1.63.
The primary differences I found when compared to the previous tests were:
- The new CalMan profiling workflow appears to suffer from elevated noise that degrades the LUT performance.
- The LightSpace grid based profiles have become significantly more efficient at reducing perceptual errors.
- DCG provides an automatic and more user-friendly method of using ArgyllCMS for LUT creation.
- The ArgyllCMS drift correction has a measurable benefit to LUT performance on a Plasma display.