Gigapixel AI test - 24MP enlarged vs 50MP native

Hi Gang,

I've been tinkering lately with some enlargement workflows in Gigapixel AI/Photoshop. I think I've finally got it dialed in where I'm pleased with the results. If anybody wants to take the Pepsi challenge, below is a blind image test for your amusement.

The test:

I took the same image below with two cameras. Canon 5Ds (50MP) and a Canon R6II (24MP). Same lens for both (35mm prime at f5.6). Taken in RAW.

I enlarged the R6II image to match the resolution of the 5Ds. Below are crops from both images side-by-side.

Enlargement method:

Export RAW file to a 8 bit TIFF. No sharpening or NR applied

Enlarge 2x in Gigapixel AI. Standard Mode, All additional settings at 0.

In Photoshop, apply a Gaussian Blur of 0.5 to 0.7 pixels to globally soften the image and reduce the "oversharp" areas that Gigapixel often introduces.

Reduce resolution with "Preserve Details 2.0" in Photoshop. This example was reduced from 12000x8000 to approximately match the 5Ds.

Apply Unsharp Mask to tastes for final sharpening.

The images:

Here's the final results side by side. Can you pick which image was generated by which camera? The 5Ds is still better (as expected), but dang...it's really close.

I'm interested in comments on why you would say which image is which?

The whole image:



And the side-by-side:


Which side was produced by which method? One side is an upres'd 24MP image, the other a native 5Ds image (50MP)



--
My site:

Having seen Gigapixel AI so highly rated I downloaded a trial yesterday and compared it to other upscaling methods of raw images including Adobe ACR enhance, PS tools, On1AI resize and the Luminar Neo add-on. I was bored on a Friday afternoon.

I'm too lazy to publish all the images, but the raw originals were sections of 20mp m43 and 24mp full frame.

I admit upfront I am not fond of the Gigapixel AI GUI but its usable.

For me none of the upscalers are universally better than the others. On some images, all at default settings, Gigapixel gave excellent results, on a few it was not so hot where others yielded visibly better results and versa vice. If you have the programs open on different monitors you can pixel peep comparisons, my rigorously scientific and objective methodology.

With all the programs the more you use it and tweak settings the better the results.

So I didn't plunk down my credit card but Gigapixel AI seems as good a choice as any for upscaling when needed. To me, as a standalone, it seems kind of pricey for what it does, but so does everything these days.

gipper51 said:

Enlargement method:

Export RAW file to a 8 bit TIFF. No sharpening or NR applied

Enlarge 2x in Gigapixel AI. Standard Mode, All additional settings at 0.

In Photoshop, apply a Gaussian Blur of 0.5 to 0.7 pixels to globally soften the image and reduce the "oversharp" areas that Gigapixel often introduces.

Click to expand...

I appreciate the workflow though adding gaussian blur biases the test.

gipper51 said:

Reduce resolution with "Preserve Details 2.0" in Photoshop. This example was reduced from 12000x8000 to approximately match the 5Ds.

Click to expand...

Why did you reduce the resolution instead of simply increasing the image size from the R6 to match the 5ds?

gipper51 said:

Apply Unsharp Mask to tastes for final sharpening.

Click to expand...

Introducing yet another random variable.

gipper51 said:

The images:

Here's the final results side by side. Can you pick which image was generated by which camera? The 5Ds is still better (as expected), but dang...it's really close.

Click to expand...

I appreciate the time and effort involved in trying to make this comparison though I think better representative images could have been chosen to make more definitive conclusions.

Adam2 said:

Member said:

Enlargement method:

Export RAW file to a 8 bit TIFF. No sharpening or NR applied

Enlarge 2x in Gigapixel AI. Standard Mode, All additional settings at 0.

In Photoshop, apply a Gaussian Blur of 0.5 to 0.7 pixels to globally soften the image and reduce the "oversharp" areas that Gigapixel often introduces.

Click to expand...

I appreciate the workflow though adding gaussian blur biases the test.

Click to expand...

I add the blur as I don't like the "overly sharp" patches that are randomly scattered throughout the image. Bias? Sure This is simply the workflow method I've chosen to apply. I'm processing these with the intention to print large and I feel this gives the best results to make a clean image. I'm just showing what I'm getting with the best I can produce.

Member said:

Member said:

Reduce resolution with "Preserve Details 2.0" in Photoshop. This example was reduced from 12000x8000 to approximately match the 5Ds.

Click to expand...

Why did you reduce the resolution instead of simply increasing the image size from the R6 to match the 5ds?

Click to expand...

I did enlarge the R6. The 2x enlargement in Gigapixel produced a 12,000x8,000 pixel image. I reduced that file to 50MP in Photoshop to match the 5Ds. You're looking at crops from a pair of 50MP images.

Member said:

Member said:

Apply Unsharp Mask to tastes for final sharpening.

Click to expand...

Introducing yet another random variable.

Click to expand...

There's tons of random variables. Everybody will have their own workflow. I sharpened both images to taste (it's mild). My intent wasn't to show what happens to each image if I apply the same settings, but rather how I can massage a 24MP image get it to match a 50MP native image.

Member said:

Member said:

The images:

Here's the final results side by side. Can you pick which image was generated by which camera? The 5Ds is still better (as expected), but dang...it's really close.

Click to expand...

I appreciate the time and effort involved in trying to make this comparison though I think better representative images could have been chosen to make more definitive conclusions.

Click to expand...

Well, can you conclude which image came from which camera? If it's that close I'd say it's conclusive that this upres workflow is doing a decent job.

--
My site:
http://www.gipperich-photography.com

bmoag said:

Having seen Gigapixel AI so highly rated I downloaded a trial yesterday and compared it to other upscaling methods of raw images including Adobe ACR enhance, PS tools, On1AI resize and the Luminar Neo add-on. I was bored on a Friday afternoon.

I'm too lazy to publish all the images, but the raw originals were sections of 20mp m43 and 24mp full frame.

I admit upfront I am not fond of the Gigapixel AI GUI but its usable.

For me none of the upscalers are universally better than the others. On some images, all at default settings, Gigapixel gave excellent results, on a few it was not so hot where others yielded visibly better results and versa vice. If you have the programs open on different monitors you can pixel peep comparisons, my rigorously scientific and objective methodology.

With all the programs the more you use it and tweak settings the better the results.

So I didn't plunk down my credit card but Gigapixel AI seems as good a choice as any for upscaling when needed. To me, as a standalone, it seems kind of pricey for what it does, but so does everything these days.

Click to expand...

Yes, none of these software packages are perfect and they certainly don't perform any magic. I've tried Adobe Super Res also and the results are similar. Gigapixel gives some more options.

My own opinion: Don't rely on this software to "invent" detail that isn't there. What it does a good job with is clean sharpening of edges giving the impression of "sharpness" for printing. It does a fine job of smoothing out the jaggies and giving you a sharper edge.

I'd almost call it "bayer interpolation enhancement"

It's certainly not going to turn a FF camera into a medium format camera.

--
My site:
http://www.gipperich-photography.com

I think the right side is the upscaled version, but its doing a great job. You should try the very latest version of topaz photo ai. It seems to have been improved with this weeks update.

Works great on old compacts. Heres a 12mp photo and an upscaled version from 12mp to circa 22mp.. hard to see any quality loss imo.


12mp


22mp

Adam2 said:

Member said:

Enlargement method:

Export RAW file to a 8 bit TIFF. No sharpening or NR applied

Enlarge 2x in Gigapixel AI. Standard Mode, All additional settings at 0.

In Photoshop, apply a Gaussian Blur of 0.5 to 0.7 pixels to globally soften the image and reduce the "oversharp" areas that Gigapixel often introduces.

Click to expand...

I appreciate the workflow though adding gaussian blur biases the test.

Member said:

Reduce resolution with "Preserve Details 2.0" in Photoshop. This example was reduced from 12000x8000 to approximately match the 5Ds.

Click to expand...

Why did you reduce the resolution instead of simply increasing the image size from the R6 to match the 5ds?

Click to expand...

I think the smallest resolution increase is 2X linear, which means you turn a 24MP image into a 96MP image. You have to then reduce it to match the 50MP of the 5Ds. That's also how it works with Photoshop (and Lightroom) super enhance. You have to double linear resolution. You can't turn a 24MP image into a 50MP image in one step. I think Topaz works the same way. I have Topaz AI, which includes the upscaling software. But I haven't tried to upscale by less than 2X linear.

Member said:

Member said:

Apply Unsharp Mask to tastes for final sharpening.

Click to expand...

Introducing yet another random variable.

Member said:

The images:

Here's the final results side by side. Can you pick which image was generated by which camera? The 5Ds is still better (as expected), but dang...it's really close.

Click to expand...

I appreciate the time and effort involved in trying to make this comparison though I think better representative images could have been chosen to make more definitive conclusions.

Click to expand...

Topaz has a post showing 2 new uprez algorithms (Standard and High Fidelity) in Photo AI 1.2.0 that they say are much better than before or what is in Gigapixel AI. Examples are included.


We’re pretty excited about the performance of these models. We’ve already been tweaking our existing upscaling algorithms in TPAI and Gigapixel for the past few years, but this new approach is truly a step above.

After these get incorporated in a new version of Gigapixel AI then possibly you will get even better results.

maltmoose said:

I think the right side is the upscaled version, but its doing a great job. You should try the very latest version of topaz photo ai. It seems to have been improved with this weeks update.

Click to expand...

I agree with maltmoose, mostly on the basis of false colour on the right hand image.

maltmoose said:

Works great on old compacts. Heres a 12mp photo and an upscaled version from 12mp to circa 22mp.. hard to see any quality loss imo.

Click to expand...

It did pretty well with that image. I use it on my old 6MP/10MP dSLR images. It doesn't do a great job with all of them, but when it works well, it's a definite improvement.

To be honest I notice more improvement on the NR side, DxO DeepPRIME/XD and Topaz DeNoise AI really help with the few high ISO images I took with my 10D. I don't hesitate to shoot my A7R4 at ISO6400 if the need arises now. Low light/moving subjects.

maltmoose said:

I think the right side is the upscaled version, but its doing a great job. You should try the very latest version of topaz photo ai. It seems to have been improved with this weeks update.

Works great on old compacts. Heres a 12mp photo and an upscaled version from 12mp to circa 22mp.. hard to see any quality loss imo.

Click to expand...

Yes, your software did a solid job on that image. I may have to give the new Photo AI a try. I'm still on Gigapixel 6.1. I'm doing the enlargement tests more as experiments and for fun, so not sure I want to pony up the $$ for new software just to play around with. But if it's that good, hey maybe.

And there's one vote for the right image being the upres'd image.

Lan said:

maltmoose said:

I think the right side is the upscaled version, but its doing a great job. You should try the very latest version of topaz photo ai. It seems to have been improved with this weeks update.

Click to expand...

I agree with maltmoose, mostly on the basis of false colour on the right hand image.

Click to expand...

Vote number two for the right side being enlarged.

You're both correct.

Alastair Norcross said:

Adam2 said:

Alastair Norcross said:

Enlargement method:

Export RAW file to a 8 bit TIFF. No sharpening or NR applied

Enlarge 2x in Gigapixel AI. Standard Mode, All additional settings at 0.

In Photoshop, apply a Gaussian Blur of 0.5 to 0.7 pixels to globally soften the image and reduce the "oversharp" areas that Gigapixel often introduces.

Click to expand...

I appreciate the workflow though adding gaussian blur biases the test.

Alastair Norcross said:

Reduce resolution with "Preserve Details 2.0" in Photoshop. This example was reduced from 12000x8000 to approximately match the 5Ds.

Click to expand...

Why did you reduce the resolution instead of simply increasing the image size from the R6 to match the 5ds?

Click to expand...

I think the smallest resolution increase is 2X linear, which means you turn a 24MP image into a 96MP image. You have to then reduce it to match the 50MP of the 5Ds. That's also how it works with Photoshop (and Lightroom) super enhance. You have to double linear resolution. You can't turn a 24MP image into a 50MP image in one step. I think Topaz works the same way. I have Topaz AI, which includes the upscaling software. But I haven't tried to upscale by less than 2X linear.

Click to expand...

Nope. The smallest scaling is 0.5x though Gigapixel AI allows custom scaling from 0.2 to 16.84). He could have used the proper multiplier, 2.083 to achieve parity. Also, given his multiple variables, it really isn't a clean comparison. Overall, all of the scaling software GP AI, Resize AI, etc. do reasonable jobs with some exceptions.

Alastair Norcross said:

Alastair Norcross said:

Alastair Norcross said:

Apply Unsharp Mask to tastes for final sharpening.

Click to expand...

Introducing yet another random variable.

Alastair Norcross said:

The images:

Here's the final results side by side. Can you pick which image was generated by which camera? The 5Ds is still better (as expected), but dang...it's really close.

Click to expand...

I appreciate the time and effort involved in trying to make this comparison though I think better representative images could have been chosen to make more definitive conclusions.

Click to expand...

Click to expand...

bakubo said:

Topaz has a post showing 2 new uprez algorithms (Standard and High Fidelity) in Photo AI 1.2.0 that they say are much better than before or what is in Gigapixel AI. Examples are included.

https://community.topazlabs.com/t/feb-2023-more-pixels-more-precision/39666

We’re pretty excited about the performance of these models. We’ve already been tweaking our existing upscaling algorithms in TPAI and Gigapixel for the past few years, but this new approach is truly a step above.

After these get incorporated in a new version of Gigapixel AI then possibly you will get even better results.

Click to expand...

Gigapixel AI has a newer update. I don't think that's the point as all of the resize software is getting pretty good.

What version of Gigapixel AI did you use? The current 6.3.3?

Your "test" included too many uncontrolled variables with a dose of subjective sharpening thrown in. In an effort to provide a more legitimate comparison - and one should note that it is difficult to provide a result without bias - I proceeded to perform a test minimizing the uncontrolled variables.

Image

DPReview's studio scene (standardized, controlled lighting)

Cameras

Canon Eos R6 (20.1k effective MP, 5472x3648)

Canon Eos R5 (45k effective MP, 8192x5464)

Methodology

Topaz Gigapixel AI v 6.3.3

R6 .jpg image imported and enlarged 1.5x (8208x3648)

Standard AI model, 0 suppress noise, 0 remove blur, 0 fix compression.

File saved as HQ .jpg

Results

See below and make the judgements for yourself. As I mentioned earlier some bias exists in this experiment (not possible to make the enlargement of the R6 image identical to the R5 in terms of size - actual multiplier is 1.49707...) and the R6 image undergoes one more compression in being saved as a .jpg. The compression algorithm used in GP AI is not user controlled (beyond 1-10) and results in a smaller test file than the original R5 file. Nonetheless, I suspect that this test will allow a better evaluation of Topaz GP AI's upscaling ability and afford the user the ability to assess resolution, artifact, etc. Obviously, one could tweak the AI model and settings to achieve a better (or worse) outcome.


R5 Studio Test Image ISO 100


R6 Studio Test Image ISO 100 enlarged 1.5x

Just saw this video about Gigapixel AI 6.3.3.

Gigapixel AI from Topaz just got BETTER

bakubo said:

What version of Gigapixel AI did you use? The current 6.3.3?

Click to expand...

I'm on 6.1. That's the latest I can upgrade to without paying for another version.

Adam2 said:

Your "test" included too many uncontrolled variables with a dose of subjective sharpening thrown in. In an effort to provide a more legitimate comparison - and one should note that it is difficult to provide a result without bias - I proceeded to perform a test minimizing the uncontrolled variables.

Image

DPReview's studio scene (standardized, controlled lighting)

Cameras

Canon Eos R6 (20.1k effective MP, 5472x3648)

Canon Eos R5 (45k effective MP, 8192x5464)

Methodology

Topaz Gigapixel AI v 6.3.3

R6 .jpg image imported and enlarged 1.5x (8208x3648)

Standard AI model, 0 suppress noise, 0 remove blur, 0 fix compression.

File saved as HQ .jpg

Results

See below and make the judgements for yourself. As I mentioned earlier some bias exists in this experiment (not possible to make the enlargement of the R6 image identical to the R5 in terms of size - actual multiplier is 1.49707...) and the R6 image undergoes one more compression in being saved as a .jpg. The compression algorithm used in GP AI is not user controlled (beyond 1-10) and results in a smaller test file than the original R5 file. Nonetheless, I suspect that this test will allow a better evaluation of Topaz GP AI's upscaling ability and afford the user the ability to assess resolution, artifact, etc. Obviously, one could tweak the AI model and settings to achieve a better (or worse) outcome.

Click to expand...

The "test" may have had more variables than preferred, but it wasn't a test of just what Gigapixel can do on it's own. It was simply showing what my the entire workflow from start to finish is producing, with one particular image. I'll admit it's biased. I showed what I want in the end image. I'm not particularly interested in what steps A & B will do on their own if the entire workflow has steps A through E, if that makes sense?

It may not be a shining example of the scientific method, but what you got was how I'd process it with what I know if I were sending it out for large print today.

Thanks for the DPR comparison. The R6 is holding up pretty well here, though there are areas you can see the R5 is pulling ahead. Not as drastic as the mere "20 vs 45" numbers would suggest, though. And it would take a nice size print for the differences to be obvious to anyone besides picky photographers

I tried my workflow on the R6 studio scene, and I think the version you posted looks better. Gigapixel 6.3.3 is doing a better job overall than 6.1. I can really see the difference in false color on the black text on white backgrounds. With 6.1 I can't seem to get rid of the false color tweaking any of the settings. I processed the RAW R6 file though and not the JPG. I get some false color on the text in both DxO and ACR with this RAW file. Gigapixel 6.1 seems to exacerbate it.

Here's the R6 upres'd to match the R5 with my workflow using 6.1







--
My site:
http://www.gipperich-photography.com