Bird AlbumHow the stickers are made
All the bird stickers on Bird Album are AI-generated. This page explains why we made them this way, how, whose photos we used as reference, and where they can go wrong.
From checklist to sticker
- 1ChecklistWe follow the AviList 2026 world checklist: 11,011 species plus 91 distinctive subspecies, one sticker each, 11,102 in all.
- 2Write down what the bird looks likeEach species gets an identification card: its key field marks, bill, legs and feet, and body proportions, in concrete colours and ratios such as "tail about 40% of total length". Errors found in review are written back into the card.
- 3Find reference photosWe pick openly licensed photos from iNaturalist and Wikimedia Commons, by several observers where possible, and put them side by side on one reference board. Every photographer is credited on the species page.
- 4GenerateAn OpenAI image model draws the bird from the identification card and the bird's appearance in the reference photos. We set the pose for every species: side-on, facing left, both feet on the ground. The photos' composition and background are not copied.
- 5Cut outThe bird is cut out automatically onto a transparent background.
- 6Strict reviewA different AI model compares the drawing with the reference photos and the card, and scores five things: identification, plumage, bill and eye, legs and feet, and proportions. A sticker passes only with full marks for identification, at least a pass on the rest, and quality good enough for a field guide.
- ↺Did not passIt goes back to step 2 with the reviewer's notes, and the card is fixed before the next attempt. If it still fails after several rounds, we find new photos and look into that species case by case.
- 7PublishEach image file is labelled as AI-generated in its metadata and released under CC BY 4.0. The review score, pass or fail, and reference sources of every sticker are public.
- 8Your feedbackSpot a mistake? Tell us on GitHub and we will generate it again.
Questions
What is this project?
We want every bird species in the world to have a clear illustration in one consistent style that anyone can use for free. Until now, complete sets of the world's birds existed mostly in printed field guides and paid databases. The stickers were first made for the birding app Mingji; they are now all open source on GitHub, and both this site and the app use them.
Why AI, and not just photos?
Photos of more than ten thousand species differ in angle, background and light, and come under many different licenses, many of which forbid commercial use or changes. Generating every bird to one standard gives each species a picture with the same angle, the same style and a clear license.
Then why use reference photos at all?
From a written description alone, AI often gets the curve of the bill, the eye colour or the number of toes wrong. What a bird looks like is a fact of nature, and real photos are the most reliable way to check it. We use the bird's appearance in a photo, not the photo's composition, light or background.
Which photos do you use?
Photos that photographers have published under open licenses on iNaturalist, Wikimedia Commons and similar sites. We prefer CC0 or public domain first, then CC BY or CC BY-SA, then CC BY-NC, and never use photos marked "all rights reserved". Each species page names the photographers and links to the photos, and the full record is in the open-source repository's data. For a few rare birds no openly licensed photo can be found; their stickers are generated from a written description, and the review checks them against published photos.
What role do the photos play?
There are two groups. Early stickers, generated before November 2026, used one photo per species, and both the bird's appearance and its pose in that photo served as reference. Since November 2026 we use photos by several photographers for each species, only to check the bird's appearance, and draw every bird in our own standard pose. The early stickers are being generated again this way, batch by batch. The note on each species page says which group a sticker belongs to.
How accurate are they?
Every sticker goes through a strict review by a second AI model, and its score and pass or fail are public. But the review is also done by AI; no ornithologist has checked them one by one, so details can be wrong. For identification, rely on a proper field guide. Stickers made without a reference photo, or that did not pass the strict review, carry an extra notice on their species page.
I found a mistake. What should I do?
Tell us on GitHub. Every species page has a link to report a mistake, with the bird's name filled in. Say what is wrong and how you know, and we will generate it again and note the fix in the changelog.
I'm a photographer and don't want my photos used as reference.
Open an issue on GitHub with the bird's name or the photo's link. Within 7 days we will generate the sticker again from other photos, and we will not use your photos in future.
How can I use the stickers?
They are licensed CC BY 4.0: commercial use, changes and redistribution are all fine, as long as you give credit.
How can people tell these images are AI-generated?
The site and the app say so. Each image file also carries the standard IPTC "AI-generated" metadata, so image tools and search engines can still recognise it after the file is downloaded and shared.