Use of AI upscaling

Another concern of AI usage on iNaturalist should maybe be AI upscalers.

Whilst it might not matter in the example below (?) …I wonder what incorrect detail will be added to images through these tools and how much other users are already utilising?

This is with :

Here it seems key features remain correct.
But, for example, the number of wing feathers appears to change considerably!

In more complex taxa, I wonder how this will impact crucial details.
Perhaps users who use upscalers should be encouraged to note for record / provide original.

Should I upload the upscaled one or the original in this instance?
I am certainly tempted to upload the latter.

I uploaded them both for now…with a note and link to this topic.


Good point. I’m not sure whether this has already been done somewhere, but now is probably a good time for iNaturalist to prominently post their policy on AI use (or come up with one if there isn’t one already) under FAQs or their policy pages or whatever. More and more questions will come up about these issues.


I also have concerns about this. If the AI is upscaling based on similar images it has been trained on, it is likely to fill in details from the most commonly photographed specimens. For organisms like insects, where hair placement/patterns or wing venation may be crucial for ID, I think there’s a reasonable probability that AI enhanced pictures could be deceptive/misIDed. If I were writing a scientific paper, I certainly would not want to base any of my data analysis on non-reproducible, AI-created photos.

AI-enhanced photos might be fine for social media or display (though I think that should still be noted), but I don’t think that they should be posted on iNat myself.


My rule for photo or sound editing has been that playing with the data collected by the camera is OK (the camera uses the data selectively in rendering the image, anyway) but adding and subtracting information is not.

By that standard, an AI rendered image such as this would not be OK. I would also think that any program that alters an image in ways that the photographer might not be able to describe fully should not be posted.

The caveat is that I don’t think I’ve thought about AI enough (beyond being reminded of the Hitchkiker’s Guide to the Galaxy description of Earth being so backward that its inhabitants think digital watches are cool, every time I read another breathless article on the subject) to have this fully figured out.


Playing devil’s advocate a little but…

Whilst I agree in theory to some extent…in practice this becomes really ambiguous… I rarely take time to use Lightroom, but AI sharpening tools are inbuilt into it now, right? …doubtless “AI enhanced photos” more broadly speaking are already in widespread usage

Similarly I guess any sharpening tool could be said to add/subtract
…and iPhone photos probably sharpen by default to some extent


The issue with AI is that it can alter the aspect of an image, inventing details that were not there. This is for AI removal of objects (a branch in front of a bird’s head etc). That should not be accepted. However, AI Noise reduction will not add detail but merely reduces digital noise. If AI is used to enhance the quality of the image without altering the content (subject) of the image, it could be allowed. For me, no AI image is ever posted for ID etc. Posting AI-enhanced images on Social media is simply a matter of taste without consequences.
Using phones with built-in AI like iPhone and Galaxy S23 etc should be avoided. Those cameras don’t even announce that they enhanced the picture using AI in the first place, making the origin of the picture doubtful to say the least.


No, I think that’s absolutely a fair point. I believe Samsung and Google phone cameras both have AI enhancement baked in at this point. I do think that there’s a bit of a more fundamental difference between tools that apply a given mathematical routine/algorithm to the pixels of a picture (basic sharpening) vs. those that essentially rewrite the picture details at a more fundamental level based on a trained dataset of other images.

With the example above, basic sharpening might emphasize or deemphasize the features of individual feathers (and potentially leave artifacts), but it won’t create new ones de novo.

In another vein, with basic sharpening the process is at least reproducible, and applies the same algorithm consistently to photos. With trained AI sharpening, the AI model is often changed/tweaked without user’s knowing and the process is much more “black box”. The outputs are also potentially dependent on other, non-local elements of the picture.


I just encountered a number of images that were clearly either generated by or heavily modified by so-called “AI” … somewhat plausible-looking observations but with completely made-up features of bird plumages, legs attached at the wrong place on a hawk’s body, extra toes on a squirrel, false color on a rat, etc. Some looked almost photo-real (others less so) and most were being blithely pushed to RG by identifiers.


Yeah i think i might’ve seen what you’re talking about. some observations in new york i think. a black rat, red tailed hawk, squirrel, raccoon is what i remember. all super smooth looking lol and i thought they were AI generated but now that i see this forum post i realize maybe they’re real photos just ai upscaled.


I would say this adjusted image is already unacceptable. The tool has inserted extra feathers that do not exist in nature, such as some weird sort-of tertials.

These are now so widespread that it would seem impossible. The effect of these built-in tools is more subtle than the example given above, so I would not be as worried about them. They don’t insert entire new physical structures into an image.


I think this website is using wrong name for their product. This is not an “enhancement” of the original image, but an AI generated copy-painting based on the original photo. Topaz’s AI sharpening, noise reduction, and super-resolution generally do not generate additional detail. (Since I only use Linux now, I haven’t used Topaz for a long time. The above experience is from several years ago.)


Definitely… but although AI upscaling might not be so commonplace now,… I see there are various apps out there already available for phones. It won’t be long before this is just another editting option inbuilt on the default phone camera app I imagine…

I should note also that on, you get settings.
I used the defaults :

Screenshot 2024-06-25 at 07.53.37

So even with this there will be settings where it becomes acceptable to some not others.
It will be very difficult to create a definite line here I think of one tool being acceptable, another not…and whilst some people may see inaccuracies in details such as feathers, others just won’t look so closely I imagine. Especially on phones.

Yeah, I’m not sure how AI sharpening works vs the older sharpening tools vs AI upscalers… but I imagine this will become an increasingly grey area too, even if possible to delineate at present.

Also, there have long been examples of images uploaded on iNat with some sort of insane “sharpening” filter applied to them which turns insects into a blocky mess. These were from long before the new wave of AI tools came out ( trying to find an example of what I mean here… but I never knew what the app was people were using which did this… anyone know the ones I mean? )

1 Like

Yeah. Like I said, the issue needs some more thought, at least on my part. The other end of the scale is the fact that sketches are acceptable evidence and there doesn’t need to be a photograph at all for an observation to reach RG, as long as some supporting evidence is provided.

My own thought (at this point) is that addition or substraction of features in a photograph should have the same impact on the observation as addition or substraction of features in a sketch would have in a submission to a rare birds committee, which is rejection of the observation. The slack in the system is greater for missing features than it is for addition of wrong features.

I’m not a world class identifier but, at a minimum, I’ll be marking anything that looks AI enhanced as reviewed and moving on. Egregiously stupid instances I’ll flag and let someone with actual authority deal with it, at least until there is clear advice on how to deal with this sort of thing.


I know I’m in the minority here, but personally I really don’t think photo editing of any kind (other than super basic stuff like cropping or lightening/darkening) have any place on iNaturalist.


The AI influence on that Vermillion Flycatcher photo significantly changes the gestalt of the face and bill to look more like a corvid than that of a flycatcher. After the change it looks to me almost like the body of a flycatcher with the head of a Canada Jay photoshopped on and the colours swapped to make it look consistent. The lighting is also very weird and looks like a low quality field guide illustration or something. Given all that, it doesn’t look like a real bird and clearly has strong AI influence to me. I think I’d be appalled to come across that while identifying and mark it as no evidence of organism. There’s no way to tell whether or not it was created completely from scratch with AI.

However I’m much more familiar with the vibe/gestalt of birds than with any other taxon; if you did this with a unfamiliar species of butterfly or a dragonfly there’s a good chance I’d be fooled.


I agree with the points above about the AI altering the gestalt of the bird. The alteration to the outline of the feathers of the wing is especially worrying to me - while Vermillion Flycatcher is easy to ID, the exact relative lengths of different feathers in the wing is often a key difference for some of the more difficult-to-ID flycatchers. The AI has made the wing feathers incoherent, so at least IDers probably wouldn’t be tricked into trying to use them as an ID feature, but if the AI gets better at passing for a real photo it could be altering key field marks. Adding AI images - even if enhanced from a real photo - will muddy the pool of images that can be used to study the species (for example, by a beginner trying to learn to ID the species, or a scientist wanting to measure some subtle features in photos to look for geographic/seasonal variation, etc). I strongly encourage you to only upload the original photo, or at least make the original the first key photo that will show up in the thumbnail.


I’ve been thinking about this for a while, along the lines of a DQA that’s something like “Evidence accurately depicts the organism and surroundings” (will need to be wordsmithed for sure), as what we’re concerned about is evaluating the evidence, not in diagnosing exactly which tools were used on the image/sound, and as others have pointed out most of these features are baked in when it comes to smartphones. This could cover generative AI, sky replacement, etc.

Many wildlife photo contests have rules that I think are useful starting points to think about this, eg

All photographs should accurately reflect the subject matter and the scene as it appeared. Photos that have been digitally altered beyond standard optimization will be disqualified. Acceptable changes include adjustments to color, contrast, brightness and sharpness; removal of dust and scratches; cropping; black-and-white conversions; and use of HDR (High Dynamic Range), where multiple exposures of the same scene are combined for a greater tonal range, and similar processes for extended depth of field. Such modifications must be disclosed. Changes that are not acceptable include photo composites (combination of two or more photos, not of the same scene); the addition, duplication, deletion or moving of objects in the photos; or the use of artistic digital filters and effects. Examples of artistic filters and effects are watercolor, neon glow, posterizing, stained glass, and others which do not show the scene as it occurred in nature.

FWIW I use DXO’s Prime denoiser and have used Adobe Lightroom’s AI Denoise feature, and haven’t noticed any issues.


not quite what you said, but.
Social media has consequences. That selfie ‘swimming with dolphins’ has consequences for the dolphins when everyone wants their selfie to go one better. By ‘upscaling with AI’


Let’s keep this discussion focused on AI upscaling (and similar manipulations) and iNaturalist, and not bringing the larger subject of social media. Thanks.


This could become yet more gatekeeping: “We only allow iNatters who have specifically looked for a mobile device that doesn’t have built-in AI.”