Topaz Gigapixel AI for enlarging heavy crops, getting better detail on birds (2 Viewers)

[Zackiedawg]

I have been using Topaz's Denoise and Sharpen for a while, and both are good tools for cleaning up shots especially for birders, when we have to go to very high ISOs or deal with a little motion blur. I recently picked up Gigapixel AI to see if it could do what I hoped, which is take shots that were fairly heavy crops, therefore low resolution and/or missing fine details, and try up-rezzing them with an intelligent enlargement tool. Normal enlargement doesn't create new pixels to build up a shot - they just enlarge the pixels there, or duplicate them. The newer uprez programs like Gigapixel claim to do a more intelligent job at predicting and calculating those additional pixels in the enlargement so that you can get more resolution that's close to the original pixels - and can be resized back down to reasonable sizes for nice detail and sharpness improvements, removing noise, and even processing as if it were an original photo - brighten, white balance changes, etc.

My typical workflow was to take shots that involved heavy crops - where the bird was a small part of the frame, and see if I could get a photo that had a reasonable resolution for display or print - at least in the realm of 6-8MP...many of these after the original crop would have been down to 3-4MP. I used Gigapixel to enlarge them 4x, then resized back down to the 6-8MP range. I also had a few shots where the final crop might have been around 6MP, but were a little soft or lacking in fine detail...I used the same process, enlarging to 4x and resizing back down. In all these cases, I was working with JPG originals, not RAW. The results have been impressive. Some examples (I have resized all of the shots for posting online, so the originals were all in the 6-8MP I mentioned - I resized them to no more than 1800 pixels on the long side for this thread):

Original of a blue grosbeak - fairly far away and simple crop yielded low resolution and not much detail:

Gigapixel 4x, then cropped and touched up:

Black-whiskered vireo - after original crop, resolution was OK, but overall shot was still soft, even after applying Unsharp mask...more sharpening was resulting in artifacts and halos, so it wasn't worth pushing too much more:

Took the same crop above, Gigapixel applied at 4x, cropped a bit tighter and resized for end result with more resolution, plus better details in the feathers and patterns:

American robin...cropped...OK, but still not crisp on the details:

Run through Gigapixel at 4x, resized back down, a little touch up on contrast, tighter crop, more details and sharpness, and more resolution than the original crop:

A very rare sighting for me in my area of Florida - a vermillion flycatcher...despite good gear and lens and a 24MP APS-C sensor, the bird was extremely distant and I had to crop very heavily just to get this at around 3MP - it fills the frame well enough, but I couldn't sharpen it too much more without lots of artifacts - and some areas in the red feathers and around the eye just didn't have enough detail to sharpen:

Gigapixel, 4x, then resized - I kept the original crop, but was able to increase the final photo to 7.2MP, bring up the shadows a bit (and have actual detail in them instead of muddy shadow noise) and still get better details and sharpness in the bird:

Just wanted to share a few examples in case any other birders are dealing with having to crop a lot, and ending up with small resolution, low details, or soft shots that they want to not just resize larger, but have better details. It's an interesting tool (I am not affiliated with the company and won't pitch any links to get me reward dollars for purchasing). Even for shots taken with P&S cameras and small sensors - it works - this shot was taken with a P&S camera in 2005, a 5MP small sensor camera with a 5x zoom:

I ran through Topaz 4x to up-rez to 20MP, then resized back down to 8MP - the end result had a tighter crop, more resolution and a larger print size, plus better details and sharpness:

Happy to answer any questions, or provide any details on the shots or samples above. Just wanted to provide a little info for any bird photographers who might not know about the software. They do have full-function 30-day trial downloads available on their site if you wanted to experiment with it.

 

[njlarsen]

I wonder why you chose to increase size beyond the level you expected to end at and then downsize afterwards. What if you had done a 2x gigapixel? is that not an option?

Niels

 

[opisska]

This kind of "AI resampling" is perfectly fine if aesthetics is your goal. But it should be stated clear that this is NOT OK for identification purposes. The whole point of the thing is literally making stuff up! So if you want a nicer picture, then it's definitely the way to go - if it's working well, it's far preferable to the other option, which is getting closer to the bird, since it's far less disturbing to the bird But a photo processed like that should never be used for an ID and should never be presented as a proof of an observation.

 

[Zackiedawg]

I wonder why you chose to increase size beyond the level you expected to end at and then downsize afterwards. What if you had done a 2x gigapixel? is that not an option?

Niels

In experimenting, I found that larger resizes, then downsizing back, actually cleaned up the noise better and seemed to do a little better overall with the sharpness/detail than the smaller resizes. It would depend on how much of a crop it was I guess - I've only been playing around with it for a few weeks now, so still a way to go in experimenting and fine-tuning for best results.

This kind of "AI resampling" is perfectly fine if aesthetics is your goal. But it should be stated clear that this is NOT OK for identification purposes. The whole point of the thing is literally making stuff up! So if you want a nicer picture, then it's definitely the way to go - if it's working well, it's far preferable to the other option, which is getting closer to the bird, since it's far less disturbing to the bird But a photo processed like that should never be used for an ID and should never be presented as a proof of an observation.

I agree in principal with that - certainly a shot which is a tiny speck of a bird barely even visible, then enlarged with software creating pixels that weren't there, can badly skew or create patterns or colors that aren't there and shouldn't be relied on for proper IDs.

However, I'd also point to the other side in the above use-cases, that the originals before processing were still clear enough and large enough for ID purposes, so that original could be used for the ID credit or proof, and then the more presentable version used for display. Basically, I think the original before any processing has to be clear enough for a confirmed ID in order for any version of the photo to be presented as an ID proof. I think you'd agree in all of the originals above, the ID is quite clear without any processing work done.

BTW - I did try it out on some much worse photos - really bad environmental conditions, terrible lighting, extreme distances, and crops of more than 100%...and yes, it did improve the cropped results...but those would definitely not be reliable for ID purposes, as the original photos were already borderline for IDs...the software can't create accurate pixels of detail that isn't there in at least some vestigial way!

 

[Victor Vector]

Nice results....I tried the free month trial of topaz but couldnt even get it to launch/open up on my laptop.

 

[Victor Vector]

This kind of "AI resampling" is perfectly fine if aesthetics is your goal. But it should be stated clear that this is NOT OK for identification purposes. The whole point of the thing is literally making stuff up! So if you want a nicer picture, then it's definitely the way to go - if it's working well, it's far preferable to the other option, which is getting closer to the bird, since it's far less disturbing to the bird But a photo processed like that should never be used for an ID and should never be presented as a proof of an observation.

I dont get this reasoning.....According to this logic NO PHOTO other than completely raw file images should ever be used to ID or proof of observation. Adding saturation/shadows/brightness et all does just the same thing, distorting the actual raw photo. I disagree and IMO see no problem with using a processed capture as ID or proof of observation.

 

[Zackiedawg]

I dont get this reasoning.....According to this logic NO PHOTO other than completely raw file images should ever be used to ID or proof of observation. Adding saturation/shadows/brightness et all does just the same thing, distorting the actual raw photo. I disagree and IMO see no problem with using a processed capture as ID or proof of observation.

The one part of it I can agree with at least with something like Gigapixel AI is if the original image is so tiny or grainy that an ID is not possible...say for example, a small non-breeding female warbler that could be any of 8 species. When you go really big on the uprez, the program is adding as many as 35 pixels for every 1 original pixel - all created, intelligently or not, from information that was never there. So it's possible that the algorithm may 'think' there should be a stripe along the eye or detects a bright sunlight patch as a wing bar that isn't really there - the created photo might actually end up looking like a perfect match for an ID, but may not actually be that bird at all. So I'd avoid letting resizing AI programs be relied on for ID purposes in any case where the ID wasn't already possible on the original.

 

[opisska]

I dont get this reasoning.....According to this logic NO PHOTO other than completely raw file images should ever be used to ID or proof of observation. Adding saturation/shadows/brightness et all does just the same thing, distorting the actual raw photo. I disagree and IMO see no problem with using a processed capture as ID or proof of observation.

That is not true. This kind of processing that you are talking about here can literally add features to the image. Otherwise it would be pointless. The "AI resample" cannot work in any other way than presuming what information is missing from the image based on its "experience" with other images. The methods that you mention are only presenting the already existing information in different ways - and that's a good thing actually, because the "raw" image is not the "most real" by no means - it's just one specific distortion of the reality that is unique in being the first one that was done with the real signal, but it's still a combination of optical tramission, sensor properties, ADC linearity etc...

 

[Musoman]

There's a review on Giga AI from pro photographer Andy Bell

AI Gigapixel Review - worth the cost? - Andy Bell Photography

AI Gigapixel Review - can Topaz Lab's latest tool, an advanced image upsizer, compete or even beat the competition? Find out here...

www.andybellphotography.com

 

[MKinHK]

Many thanks Zackiedawg for a really useful original post. Especially appreciate you explaining your workflow AND posting the "before and after" results so that we could see the improvement for ourselves. The American Robin shot is particularly striking.

Could I ask what camera you were using for the original shots?

Cheers
Mike

 

[Zackiedawg]

I've shot with a variety of Sony gear over the years - in film days, I was Pentax and Canon, but I became familiar with Sony digital in 1997 and kind of grew into them, so once they started doing DSLRs and other interchangeable lens cameras, I ended up staying with them - first with DSLRs, then with mirrorless. The shots from the samples above span a range of cameras - the robin was with an A580 DSLR, as was the grosbeak. The vireo and flycatcher were with an A6300 mirrorless.
And thank you for the kind words...glad to know it was of some value or use!

 

[Chosun Juan]

+1 👍 I agree - an interesting thread JDWG.

I just want to chime in quickly and in general since I have no experience with the software discussed (but have seen great 'refocusing' and sharpening results in Topaz).

On my Samsung S21 Ultra, which I bought for the cameras, the you beaut AI (or whatever) zooms are so riddled with artefacts as to be just about useless beyond the optical 10x. It looks like an artsy filter applied with the result being more Art than subject ......

Your results with this process are in a different league. I understand the point of concern over 'fictional' constructs creeping in though.



Chosun 🙎‍♀️

 

[MKinHK]

As long as the original is included with the enhanced image I see no problem with using enhancement tools for identification, especially if the tool used and the settings applied are also stated.

Cheers
Mike