We’ll have to wait for an official response, but having read their developer comments as this was being developed I would think that the difference between version 3 and 4 is that they turned the AI back up again, and that’s why it’s slower.
They’ve said before that they started with a very very detailed AI that they turned down before the initial release to speed things up; looking for the right balance of quality and speed. Version 4 is the first to provide the user with 2 levels of AI, and that probably just reflects their recognition that after more training of their AI engine to make it better for version 4 it got noticeably slower again.
That forced them to give us the second option of using a somewhat neutered version of the AI. It’s not just a clone of the fast one from version 3, because they still needed to fix some of the issues like the grid patterns and odd lines showing up here and there.
Going back to the beginning we had a standard image to compare speeds with, so I reran the test again to see where we’re at now.
Doing this standard 16 MB image again at 4X with Version 4.0.0 using the best AI model, max RAM, and Noise Suppression at Low and Remove Blur at High took 156 seconds. (Turning Noise and Blur off only saved half a second.)
That compares to 377 seconds in version 1.1, and 93 seconds in version 2.
Using the not best AI model in version 4 it dropped back to 74 seconds, so that method is even faster than version 2 was.
I wish I had run this test with version 3 to compare that as well, but I seem to remember that not being a significant change from version 2.
p.s. I just ran this test again with version 4.0.1 and the times are identical to 4.0.0.
Just installed and tested, and it is faster.
Not version 3 fast but certainly more workable.
Haven’t tested for image quality yet (Topaz may have just wound down the quality to get some speed into the thing).
Keep on with the good work Topaz, when version 4 is as fast (or faster?) than version 3 I will be a happy chappy.
Unfortunately no statistically significant change for me with V4.0.1 at 1m54s.
Results all at default 4X settings.
Gpxl AI V3 (1 GB Radeon HD7770): 3m33s to 3m40s
Gpxl AI V3 (6 GB GTX 1060 419.35 driver): 53s
Gpxl AI V4.0.0 (6 GB GTX 1060 419.35 driver): 1m58s
Gpxl AI V4.0.0 (6 GB GTX 1060 419.67 driver): 1m58s
Gpxl AI V4.0.1 (6 GB GTX 1060 419.67 driver): 1m54s to 2m02s
In my mind it’s takes a pretty strong conviction in programming choices to release software that only runs half as fast as the previous version or obsoletes most of your client’s hardware due to performance.
However, for me it’s not currently a super huge deal as I’m not sure how often I will scale up images and I can always be doing something else while the scaling goes on in the background.
As the models were updated to provide a better scaling and to correct issues such as errors visible in the output it is to be expected that there is more processing taking place. The proof is in the quality of the output and that is easy to compare.
For those that believe the processing is too slow there are things that you can do to tune the application to suit your processing needs as well as making sure that you meet at least the recommended technical requirements for the application. Tuning and performance tips can be found here: