Another question for Taylor:
In batch mode would it be possible for GP AI (and the other Topaz apps) to generate a sound alert once completed?
I often use Topaz apps in the background while doing other things. Sometimes I am not even at my computer while it’s running!
I agree, Ray.
If the term is “processing” then the image in progress should be counted. It keeps throwing me that it says 0/1.
If the term is “processed” then there could be a zero count. Until the image is finished.
If “processed” is used then the joke monologue is irrelevant b/c it wouldn’t be the intended distraction during processing.
This is minor but when processing with my CPU + Maximum quality the Generating preview counter goes from 0% to 1% but no higher. After 1% it finishes. With Max quality off it stays at 0 then flashes 1 or 2% before finishing. With the GPU it does count up.
Hey, a quick question that I’ve been wondering for a long time.
Gigapixel AI can go up to 6x upscaling, right? However, I’ve noticed that no additional detail is being added past 4x upscaling.
I did a quick comparison by upscaling an image at 6x, then upscaling another one at 4x and then changing the resolution of the 4x export to match the resolution of the 6x export. Then I layered them on top of each other in Affinity Photo and did an A/B comparison. Both images were identical. Is it just me and is this is normal? If this is the intended behavior, why is there even a 6x option? Thanks in advance.
It’s not clear how you did the change in resolution. My take is that you saved the 4X file then brought it into Gigapixel again and scaled it at 1.5X to match the 6X scaled picture. If that is the case then of course it is much easier to scale 6X rather than the extra steps to do 4X and 1.5X.
Sorry for not making that clear. I changed the resolution of the 4x export in Affinity Photo. A simple Bicubic resizing.
I tried that with one of my photos and I did see a small difference with the GAI 6X picture being slightly sharper than the Affinity Photo upscaling of the 4X picture. However it was only noticeable in some clothing. It is still simpler to do 6X in GAI instead of two scaling’s. If more than 6X is needed then using Affinity or Photoshop to do the final upscale will most likely work better than going beyond 6X in GAI.
Thanks for the test. So it seems that it’s not only me. In my tests I don’t see any difference between 6x upscaling vs 4x upscaling + resizing in a photo editor to match the 6x resolution. Isn’t this false/misleading advertising then? Wherever I look, all adverts for Gigapixel AI say that we can get up to 600%(6x) enlargements with detail preservation:
“Enlarge your photos up to 600% while perfectly preserving image quality.”. Taken from the main page of Gigapixel AI.
I was wondering if there was something wrong with the software and my installation. Turns out, it doesn’t actually preserve detail past 400% by design? I’ve done many tests with a variety of different images and so far this indeed seems to be the case.
I don’t think I’ve ever seen anything about this for Gigapixel, but it is understood with VEAI that the ONLY AI upscaling that happens are 2x and 4x. If you say 3x it goes to 4, then uses simple non-AI downscaling to 3x. The breaks, if I remember right, are:
lt 1.2 = 1x
1.2 - 2.4 = 2x
gt 2.4 = 4x
The scaling used by VEAI for anything but 2x and 4x is Lanczos resampling
Developer has said that they were going to migrate the AI engine in VEAI to Gigapixel, to fix the GPU artifact problem. Now V5.3.1 version, the GPU artifact is fixed, may be they have switched to VEAI AI engine already and act similar to VEAI.
I have a different conclusion to the test. If you use a 4X upscale and GAI improves the detail and even removes most of the Jpeg artifacts then it has done its job. Further upscaling in a photo program using Bicubic or other methods simply increases the number of pixels. The hard work of restoring detail is already done. If you use a 6X upscale in GAI it will add more pixels and do the corrections just like the 4X version. I expect them to look very much alike (4X GAI + Bicubic vs 6X GAI). Both times GAI is doing what it’s supposed to do.
1 Like
I’m sorry, but I don’t agree. By the same logic you can say “up to 10 000% upscaling while perfectly preserving image quality”. I still think it’s misleading that there’s a 6x button as a built-in preset. I am sure that I’m not the only one who thought that due to the 6x preset being there, GAI restored details up to 6x upscaling when in fact it only does so when doing up to 4x upscaling. I think the “6x” button should be removed to avoid misleading people that it restores details with the AI engine when in fact it doesn’t and it only interpolates the image from the 4x AI upscaled image to 6x with a simple bicubic sampler.
1 Like
For me the 6x button is a reminder that anything above that is going to be processed differently. Specifically, all the image detail is lost to blurriness at 6.01x, whereas 6x is still crisp and clean.
I can see both sides of the argument… however, I feel that if the product is advertised “6 times” or “up to six times” and there is an option in the application to scale 6 times, I would believe that the product is upscaling the full six times using the machine learning technology. Indeed, it was a revelation to me when I first learned the product actually scaled using machine learning by 4 times, then used lanczos/bicubic for the rest of the way to 6 times.
With VEAI, I was scaling like most users by inputting the output size. But when I saw the “1x” and “2x” and “4x” models in the directory, and reading the community threads, I now know that VEAI is scaling with the machine learning only by those factors. So now, I only scale by 2x or 4x in VEAI, then do final scaling (using Spline144) in Hybrid/AviSynth.
Likewise, with Gigapixel, I will only scale by 2x or 4x, then use other software to process further.
I know “6 times” is a bigger number, but when you put a 6 times option in the “machine learning scaler” application and it’s actually only scaling by 4 times then doing a regular lanczos/bicubic resize, there is some level of deception involved there.
It would probably be better to either take the options for > 4x scaling out, or make it more clear to the user that the actual machine learning scaling happens at 1x, 2x, and 4x factors, and anything else is being done using traditional resizing.
2 Likes
Where did you see this? Was it on the Topaz website (not just in a discussion)?
1 Like
I’m pretty sure discussion only, and mostly on the FB beta site. The earliest reference I can find that spells out the 1.2/2.4 breaks is a post by Matt Lathrop 10 weeks ago, and he was at the time a Topaz engineer. He states it as fact, not conjecture. his specific post reads
It’s not mush, but yes it’s a problem. Very annoying.
On my side I noted that repetitive motifs/pattern like skin (or vegetation) may do this at the end. You can see that they are in fact the border of squares (engine seems to work by blocks of pixels) that go crazy.
But there’s a “workaround” : the effect shows up when ‘remove blur’ is near or superior to ‘suppress noise’.
For example:
80/50 (noise/blur) → OK
80/70 → Meh
80/80 or 80/100 or 20/40 → Artifacts
But yes it’s a real problem.
What were your settings ?
Based on comparisons it could be possible for the scaling above 4X to be using bicubic but if that is true then why does the results above 6X get far worse? Using the bicubic method it could still produce good results to at least 8X.
The end result is that 6X scaling still is far superior to just using standard methods like Bicubic. Below is a test using the 1000 x 1000 px paintbrush picture that comes with Gigapixel. You can see that the 4X + 1.5X is slightly sharper than the 6X version but I believe that is due to extra sharpening by running twice. The 6X + unsharp mask is equivalent. All settings were reduce noise = 40 and reduce blur = 40 and picture crops are 50%.
6X bicubic only:
Gigapixel 4X resized to 6X using bicubic:
Gigapixel 6X:
Gigapixel 4X then 1.5X:
Gigapixel 6X + unsharp mask:
1 Like
My favorite Unsharp Mask settings are Amount: 20%, Radius: 30 pixels, Threshold: 0. This provides a nice touch of local contrast enhancement to make the image look crisp (less veiled) without creating any sharpening artifacts.
