AI Enhancement and observation quality?

[headsoup]

What’s everyone’s thoughts on things like Topaz being used on observation photos?

In general, it’s quite amazing the way they can upscale, denoise and sharpen images, but I don’t want to use it myself because I feel it’s potentially creating misleading images and they’re kind of ‘fake real.’ I know the accuracy is pretty good (and the photos look amazing), but it’s also filling in details through algorithms and I wonder if this has a possibility to create incorrect observations (e.g. a mis-shapen spiracle) that ID’ers won’t know of because of how well the software sharens/upscales.

Do you use it?

[cthawley]

I personally don’t use it due to the same concerns that you have. It’s creating new details based on a black box approach that we don’t understand.

If the point of the picture is just to be pretty, I don’t see an issue (though I do think it should be noted that it’s an AI-enhanced pic).

BUT if the point is to enhance IDability, then I think it’s a bad move. For instance, what if the AI is trained on different pics that include the most common species? It could fill in details based on that common species and prevent IDing as less common species that aren’t in its training set (don’t know enough about how it works to know how realistic this scenario is).

There are some other thoughts in this thread: https://forum.inaturalist.org/t/photo-sharpening-does-it-help-or-hinder-the-inat-ai/33021

[broacher]

I use tools from Topaz. AI Sharpen and AI Denoise, specifically. Plus on occasion, Gigapixel. I also use DxO PureRaw.

It all comes down to the words ‘use’ vs. ‘over-use’. And it takes a while to learn how to get the best and least noticeable results. I’m still learning!

I find that the denoisers are likely the most useful tools. PureRaw only works on RAW, and only on supported cameras. But it works so well that I depend on it now for camera purchasing choices! The company has invested a lot of R&D time and money tweaking the RAW models in all the different cameras and lenses, and it really shows! PureRaw is more of a batch processor. You can drop a bunch of files or a folder into it and go get a coffee while it ‘cooks’.

What I have found so helpful about the denoisers (I also use Topaz Denoise, sometimes in combo with PureRaw) is that they open up the high ISO range of any of the cameras that they support. The detail that is found is real, it’s just lurking in the pixel patterns that only the RAW format conveys. When you view a high ISO shot on the camera’s screen it can look unbelievably bad and noisy. The camera cannot process to the same level that a full blown, fairly powerful, laptop or desktop with a high-end GPU can achieve. And even then, you could be looking at anywhere from a few seconds to a minute per image to get where you want to go, and see things cleared up.

In some ways, I like PureRaw’s limited adjustment options as compared to Topaz’s multi-model workflow. It’s far easier to get artifacts introduced when you have users tweaking all the sliders.

There’s a procedural sequence to the Topaz lab stuff too. Some people don’t follow that. Or don’t know enough to judge when they can modify it. Or, very commonly, they crank the sliders too much or they push the algorithms to fantasy levels.

On the other hand, done right – and suddenly, there’s stuff in the shadows that you didn’t see at all. Or tiny hairs appear in what looked just like little fuzz lines.

But all in all, I embrace this tech for what it can achieve in terms of species identity work. I know that rankles a lot of photo purists, but as I have said many times, we’re not doing this for art. If it looks pro and beautiful – great! But more important is nailing an ID.

Right?

[pisum]

there’s really no difference between what Topaz does and what many mobile devices and cameras already do. to the extent that such tools can help you be a better observer or appreciate nature and share that appreciation, i think it’s fine to use any of those tools that are available to you.

[fffffffff]

I don’t use it because all examples people shared here looked bad, if you have a blurry photo, oversharpened photo is not a solution.

[headsoup]

Can’t say I exactly enjoy trying to ID from many of the mobile phone ‘enhancements’ but it’s a fair point, JPEG processing is a lower-scale implementation of what these tools do.

@broacher

I find that the denoisers are likely the most useful tools

I agree, especially liking macro. I think if there was one I would use it would be de-noising so I can up the iso and rely less on flash. But… I think I’d still do a fair few side-by-side comparisons first to see what changes between using ISO 200 and ISO 3200+denoise.

[lynkos]

The key words in all this are “intelligence” and “control”. Not just that of the AI tools, but above all that of the photographer/processor. It’s all too easy to forget that when shooting in jpg, ALL cameras apply a sort of AI to the image, and their choices may well not be the most valid. This of course changes when shooting in RAW, but a RAW image must be processed and converted to jpg for posting to the web, and again, all more-or-less automatic programmes employ some sort of AI during the conversion.
I regularly use Topaz DeNoise and it enables me to get decent IDable images at otherwise impossibly high ISOs, “capturing” life forms in conditions where otherwise I would have had to use a flash… often equally destructive of detail and/or colour and a potential source of disturbance to the life form. But although the intelligence may be artificial, it’s ME that controls it to give back an image as close as possible to what my eye and brain (the best intelligence ever) saw in real life. For example, that could mean NOT applying the denoising to the entire image. If I see that critical detail could be lost during the process, I may mask sensitive areas and apply the denoising to just some parts of the image (easy to do in Topaz) where untreated noise may in fact make other details hard to evaluate.
Although I’ve not used other AI programmes, the same reasoning would apply, but with even more need for INTELLIGENT control on the part of the user. On some photos it may work, less so on others. And in the case of a blurry and/or very small image, “therapeutic obstinacy” is rarely worth the effort. Unless the observation is particularly valuable, I find the recycle bin is by far the best option.

[tiwane]

I use DXO PhotoLab’s “PRIME” denoiser on some of my noisy photos and in general I’m happy with the results. I agree that it’s not much different than the various techniques smartphones.

[annkatrinrose]

I think minor image processing to enhance the picture quality is fine, but image manipulation that significantly changes the story or subject of the image is not. If you look at the commonly agreed-upon definition of nature/wildlife photography, it says:

“No techniques that add, relocate, replace, or remove pictorial elements except by cropping are permitted. Techniques that enhance the presentation of the photograph without changing the nature story or the pictorial content, or without altering the content of the original scene, are permitted including HDR, focus stacking and dodging/burning. Techniques that remove elements added by the camera, such as dust spots, digital noise, and film scratches, are allowed. Stitched images are not permitted. All allowed adjustments must appear natural. Color images can be converted to greyscale monochrome. Infrared images, either direct-captures or derivations, are not allowed.”

Those rules were made for photography competitions and such and not for iNat, but I think they give some good guidance for questions like this.

[jasonhernandez74]

I can’t imagine what is wrong with a direct-capture infrared image, especially if the alternative would be to disturb a nocturnal animal with a flash.

[annkatrinrose]

Yes, and I personally can’t see anything wrong with stitched panoramics for scenic landscapes either. Curiously enough, HDR, focus stacking and black-and-white conversions are expressly permitted, but infrared and stitching are not. My understanding is that this set of guidelines was agreed upon by several photography societies as a common definition for nature/wildlife photography contests, exhibitions etc. That’s a bit different from taking pictures for scientific purposes or iNat records.

It goes on to define acceptable subjects for Nature Photography (e.g. landscapes, weather, geology, zoos, botanical gardens) vs. Wildlife Photography (strictly wild organisms unrestrained in their natural habitat) as categories for photography competitions. There is some overlap with what is acceptable and sorted into wild vs. casual for iNat observations but not completely.

[broacher]

@annkatrinrose

All allowed adjustments must appear natural.

Wow. Must be a lot more super-experienced and all-knowing judges out there than I imagined.

It’s funny, as I was reading that list it brought me back to all the time I’ve spent in my corporate design career retouching the publicity photos of executives, VIPs, and happy customers to remove natural ‘flaws’ and enhance ‘appearances’ under the provision that the results look ‘natural’.