Use of AI upscaling

General
62

I’d be happy if someone could name just one piece of proprietary camera equipment that is calibrated for iNaturalist work. The reality is that no camera equipment is specifically calibrated for this platform. Cameras are designed for general use and are not tailored to any specific observational application, including iNaturalist.

I am not merely equating AI manipulations with benign photographic adjustments. My argument and intent are to broaden the understanding and scrutinize all forms of image manipulation to be consistent. By that, I mean that we should not single out AI-induced artifacts as uniquely problematic when other, more conventional photographic techniques and anomalies (e.g., lens choice, focal length, rolling shutter, etc.) also significantly alter images. While AI-generated details and traditional artifacts arise from different technological foundations, they are analogous in how they potentially distort the reality captured in images; they both impact the image’s authenticity and reliability for accurate identification and analysis on platforms like iNaturalist.

As an example, there are instances in photography where details are not added but rather omitted, such as when parts of an image are under or overexposed or simply out of focus to the extent that no information in those areas of the image is recoverable. This raises a question: if a photo can be tagged as research grade (RG) on iNaturalist even though it might have missing information due to traditional flaws, why not a photo that has extra information from generative AI?

To clarify my point, iNaturalist’s criteria for achieving Research Grade status require that observations have a date, location, photos or sounds, and are of wild organisms. Additionally, the community must agree on the identification to at least the species level with a 2/3 consensus. This ensures that the data shared with scientific partners is as accurate and reliable as possible. Moreover, iNaturalist acknowledges that poor quality photos can still be useful if key diagnostic features are visible, as demonstrated by the example of a blurry and heavily cropped photo of a Wedge-tailed Eagle that still reached Research Grade due to the diagnostic shape of the tail. This example reiterates my point: if such a photo with so much missing data can be used for identification, then why not a photo altered by generative AI that adds a few feathers to a bird’s wings but retains other distinctive morphological features and accurate location data?

An earlier post highlighted the critical difference between a drawing and an AI-generated digital image, noting that a drawing is unlikely to be mistaken for a genuine photograph. This observation raises a crucial question: Given the inherent flaws and biases introduced by cameras and lenses, what exactly does a ‘genuine photograph’ look like? We can argue that all photographs are, to some extent, interpretations of reality. Just because a ‘real image’ captures light rays on a digital sensor or film, it is still subject to the interpretation and manipulation of that light through various means. So how do we define and agree upon the authenticity and accuracy of a photograph in scientific and observational platforms like iNaturalist, or anywhere else for that matter?

1 Like
63

You don’t see the qualitative difference between say, a photo of a bird that only has one leg clearly visible in it, vs’ a photo of a bird that has 3 legs because the “artist” (whether algorithmic or human) added things to it after the fact?

This is going to be a hard conversation to have if there’s no agreement on that.

then why not a photo altered by generative AI that adds a few feathers to a bird’s wings.

In fish, the number of rays a fin has is quite often a part of the species description and can be diagnostic - I’d assume much the same is true for many birds, including details about the feathers themselves.

A 3 legged bird with too many feathers, unless a verifiable mutant, probably shouldn’t even be a candidate for Casual let alone RG status.

Things that aren’t visible in an image are known unknowns - and we know how to interpret those. Things that are “seamlessly” added to them without disclosure are unknown unknowns, and you don’t even know they are there to know what to do with them, which has the effect of distorting everything. When someone really uploads a genuine 3 legged bird, nobody is going to believe them. Genuine outlier individuals with the ‘wrong’ number of feathers or toes will be indistinguishable from “useless” reported fakes.

That’s not the same as a little lens flare or chromatic aberration, or any other artifact from camera equipment and image processing that is intended to optimise fidelity in the reproduction.

Deliberate addition of features that were not part of the original image, whether by “touching up” by a human actor, or interpolation from things not in that image by an algorithm are not Evidence Of Presence.

The difference between those two things is what I understood @spiphany to mean by “calibrated” (originally your choice of words) for this purpose. It seems like a relatively clear distinction between widely separate goals of the tools in each camp to me …

One tries to make what is present as clear as it possibly can - the other uses what is there merely as a template or muse in the process of making an Original Creation of its own. One is an observation (so “calibrated” for the use inat requires), the other is Art.

Nobody is going to mistake the images in Kunstformen der Natur as actual observations - but if they were painted instead by Vermeer, we might be having a similar conversation if someone tried to pass them off as an observation.

7 Likes
64

more than 2 thirds - so 5 against 2.

1 Like
65

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.