I myself have difficulties with my own models sometimes because some inaccuracies occur at the interface between the GPU and the CPU.
You wrote:
“… the issue itself is one of the more complex bugs to fix”
Maybe instead of digging through kilometers of code in tensorflow, you’d better try adding another model to Gigapixel that will remove “banding artifacts” from the image, according to my concept:
This will be faster than fighting entropy in the form of machine errors . and adding a trained filter shouldn’t slow down the entire engine in Gigapixel 5 too much.
Best Regards,
Lech Balcerzak
PS.
All the time I admire you for your determination and new ideas (eg bypassing the problem of not scaling images in the “Filter” menu and designing the plugin as “Automate”). Very tricky!
I certainly got the same result when I processed your original image in the latest beta. However, I tried the experiment of photographing four blue objects from around my house: two fabric, one painted wood and one shiny painted metal. The colour shift was not duplicated in any of them, so I suspect it relates to that particular shade/luminosity of blue.
You might want to raise a support ticket here:
I also noticed that ages ago (GAI also likes to add blue-ish tint at the edges of gray-scaled drawings sometimes), but it probably requires significant fight with the training model on devs side.
If you are sure that your trial version has not expired then you should contact Tech support from the main Topaz website. I noticed that in your task manager picture that Gigapixel is 3.3MB while I checked mine and it is 133 MB. So your computer is not running the full program.
Thanks for testing those! It’s an odd issue for sure, but given the older version doesn’t have it, surely something went “wrong” around v4.5.0
I’ll give the support ticket a shot, but given it was brought up back in 5.0.2 and wasn’t fixed yet… I’m not sure if they consider it a large enough issue to fix.
Thankfully, it’s not a colour that’s encountered much in the real world! The engineers often want samples to run through their AI learning machines, so offer them your’s.
Anyway, it’s easy enough to remove using any photo editing software. (I have to do this all the time with JPGs taken on my digital camera of 20 years ago, where there is almost always a cyan bias in the skies). Here’s what On1 achieved with one of your GP’ed images, by moving the purple slider toward the blue end of the spectrum:
Currently it is impossible to handle these large batches because it tries to load all the images for some reason. It should only load the image that you clicked on.
Maybe add another button. “Remove All Finished Images” or something like that. Or add a check box to allow that functionality. When you have a TON of images it would be helpful.
Something definitely happened to the models around v. 4.5. I often compare quality of image (by zooming in at pixel level) between 4.4.6 and 4.5 (and newer). Small details in the latter versions are more blocky, as opposed to finer details in 4.4.6. That difference still exists in the latest version.
Why have you removed the possibility to drop files into the window with the mouse?
It’s incredibly time consuming and annoying to have to add files via the open files dialogue, this will force me back to an older version if there’s no way around it.
Plugin mode?
I open the program from the start menu and when I then try to drop images dragged from explorer I get a forbidden sign (crossed over circle) and nothing more happens.
. and adding a trained filter shouldn’t slow down the entire engine in Gigapixel 5 too much.
