Saturday, September 8, 2018

Topaz AI Gigapixel - tested

When Topaz came out with their AI Gigapixel (AIG), image resizing product, I was sceptical. Firstly, I had tried many resizing products in the past and had settled on one that worked quite well (within the accepted constraints of resizing). Could AIG, do better than what I was already using? I doubted it.

My first attempt was not good. AIG crashed on first use and clearly didn't want to work with my graphics card. I walked away disappointed but not altogether surprised. I told myself that, anyway, it probably wouldn't have been any better than what I was already using.

I must be a glutton for punishment because when the email arrived announcing an improved 1.1 version, I downloaded it and tried again. This version included a CPU mode, which bypassed the GPU on my graphics card. That should, at least, allow it to run.

It did run, and the first output looked surprisingly good. So I devised a stricter test - I took a 24MP DSLR image and downsized it, in my normal editing program, to 6MP (half the width and height). I then put the 6MP image through AI Gigapixel at 200% to bring it back to its original size. I could then compare the original 24MP image with the enlarged 24MP AIG image and see how different they were.

The AIG image was sharper than the original. AIG obviously adds some sharpening during the resizing process. To compensate, I added some sharpening to the original file so that they were more directly comparable. What I was really looking for was a loss of detail and fat 'jaggy' edges in the enlargement. At 100% I could not see any difference between the two files. Definitely, not the result I was expecting.
Screenshot - Original 24MP file on left - 6MP file enlarged to 24MP on right (click to see full size) 
The screen-shot above is taken at 200% magnification and contains both in focus and out of focus items. The original image is on the left and the AIG resized image is on the right. I am still struggling to see any significant difference between the two images. Not quite believing what I was seeing, I repeated the same test on other images all with a similar result.

As a final test, I took an image from a different camera - a compact with a 1/2.3", 16Mp sized sensor. I enlarged the image by 200% in my image editor, using the bicubic sampling method and then enlarged the same image in AIG and compared the results. The screenshot below, is at 200% magnification and the results are quite surprising - the AIG image has much finer detail (particularly at the edges), texture seems to have been replicated rather than enlarged, and image grain has been suppressed to a far greater extent, leading to a much cleaner image that would be difficult to pick as an enlargement.
Screenshot - Bicubic enlargement on left, AIG enlargement on right (click to see full size)
All my tests with AIG have been at 200% enlargement because that is what I would normally use. But AIG can enlarge up to 600% (though the results may not be so acceptable). The shrink-enlarge experiment indicates that AIG could be an effective substitute for a camera sensor with four times as many pixels (16MP becomes 64MP) or for a longer focal length lens. AIG produces significantly better results when compared with either a standard bicubic enlargement or previous industry-leading enlargers.

US$99 does seem a lot for a single task piece of software. But when compared with the cost of a camera with a larger sensor or a lens with a longer reach, AI Gigapixel may well be a reasonably priced alternative.

1 comment:

  1. Having passed the comparison between an original image and the same image down-sized by half and then up-sized in AI Gigapixel to the original resolution, I have now down-sized to 1/4 and up-sized by 400% and also down-sized by 1/6 and up-sized by 600%.

    The 1/4 > 400% image showed some artifacts and softening when compared to the original on screen at 200%. They were difficult to spot and, when the magnification was reduced to 100%, almost imperceptible (even knowing their existence). The results would be acceptable in all but the most demanding of situations.

    The 1/6 > 600% image was noticeably softer, even at 100%. and there were the odd pieces of small image detail missing when compared with the original. Nevertheless, these differences were only detectable by comparison with the original and on its own, the 600% image would stand up well on social media and in large print applications.

    I find these results quite remarkable and far superior to any other image resizing program I have used. By the way, AI Gigapexel is *SLOW* - a 16MP image took 7 minutes to up-size by 200%, so don't bank on applying this to large batches of images unless you have a *very* fast machine (mine has an i7 processor, 2GB graphics card and 16GB of memory).