Coming soon: Deep fake nature photography?

Here are some true flies it generated with the prompt “A new species of true fly on a leaf, photograph, photorealistic, high definition, award-winning photograph” -





And yeah, it took me a few months of being on the waitlist to gain access. I know I definitely couldn’t wait either!

One thing to note, it doesn’t seem to do great with scientific names. Inputting the same scientific name can output greatly different results, from insects to mammals to pretty much any other life form. It doesn’t seem to be trained on scientific names, for the most part. So it may be difficult for someone to generate an image for certain species, especially ones without common names.

5 Likes

I think there’s a good chance of people submitting AI-generated images just out of curiousity to see what iNat’s CV says for them. I already see people innocently/naively doing this with the iNat app; submitting pictures from the internet or friends or whatever. They generally are only aware of the app’s photo-ID capabilities and don’t realize that iNat observations also contribute to citizen science databases.

Here’s another interesting example of AI-generated nature photography. You draw a low-quality version of a landscape and the AI generates a photorealistic version of it: https://www.youtube.com/watch?v=eaSTGOgO-ss

5 Likes

Thanks!

Interestingly, the flies are much weaker and more obviously generated than the butterflies… I guess the more complex and variable anatomy is not so easy to get away with. There is a little more complexity to your prompt though as well.

1 Like

The butterflies’ heads have problems but the oversaturation/excessively vibrant colours immediately make me suspicious of them before I notice that, whereas I only see problems with the flies when I look really carefully at the legs and thorax details. Familiarity with the taxon is going to be really important as these get better.

2 Likes

I would argue the first fly is pretty good, others are clearly fake.

1 Like

Maybe not like the completely different spots in the example, but it looks like we may already have pics on iNat that unaware users have added synthesized details to while fixing blur using an AI enhancement:
https://www.androidcentral.com/apps-software/samsung-galaxy-enhance-x-photo-app

Great question! @Meowdle I didn’t realise we were already at this point with DALL-E2 of being able to generate convincing images of new species, great addition!

A couple possible answers I was able to find:

  1. I found this blog post which essentially points to using ML to detect ML-generated images:

https://blog.jayway.com/2020/03/06/using-ml-to-detect-fake-face-images-created-by-ai/

This post describes 99%+ accuracy in using this idea for faces as a browser extension:

https://petapixel.com/2022/03/18/chrome-extension-can-detect-fake-profile-pictures-with-99-29-accuracy/

  1. An answer beyond faces may already be here through detection of digital artifacts which are generated in the process of generating the images:

https://www.mdpi.com/2313-433X/7/8/128

https://www.sciencedirect.com/science/article/pii/S1077314222001114?casa_token=KptT93R3X14AAAAA:5tWU9BH-nIUtIRJvhUXBzb0FO3Ij9fsYVW4OOJziTlZuwyokyjeH7kPYJa2bhWiQfBqsAXKm8vg

It seems to beat AI, we’ll have to use AI :)

  1. “Responsible” companies can (and should) be embedding artifacts which are not easily edited or detected by laypeople as a mark of their generation through ML. Similar to the printer ink phenomenon for anyone who’s already familiar with that one.
1 Like

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