huashu-bird-sticker: AI-drawn field-guide illustrations for 10,973 bird species

One illustration for nearly every bird species in AviList v2026, the unified global checklist: There are 11,064 images: one for each of 10,973 species, plus 91 subspecies. Each one is a single whole bird, mostly in side view, cut out with a white sticker border on a transparent background, in the spirit of a field-guide plate. The 3,561 images redrawn since November 2026 all use one standard pose: head to the left, wings folded, feet flat on an invisible ground line. Older images vary more, and some show the bird on a perch.

These images are AI-generated. We drew them with an image-generation model (OpenAI GPT Image), using reference photos and written descriptions of each species' field marks. Then each image was reviewed against the species and given a score. They are not photographs and they are not paintings by human illustrators. Read Limitations before you use them as ground truth for anything.

• Source repository: https://github.com/alchaincyf/huashu-bird-sticker
• Zenodo archive (DOI for all versions): https://doi.org/10.5281/zenodo.23968399 ; this version 2.2.0: https://doi.org/10.5281/zenodo.23968400
• Snapshot: data version 2.2.0 (2026-11-05). The repository is updated as stickers are redrawn; this dataset is a fixed snapshot with a DOI.
• License: stickers CC BY 4.0, data compiled by the project MIT, third-party fields keep their own terms (see NOTICE in the repository).

Why this might be useful

Field-guide illustrations are a known source of side information for recognising bird species that have no training photos (Rodríguez et al., WACV 2024). The illustrations used in that work come from Birds of the World and are subject to that project's subscription and license terms. This dataset is a freely licensed alternative with similar properties:

• Coverage: one image for 10,973 of the 11,011 species in AviList v2026 (the other 38 were recently split and are still being drawn), across all 45 orders and 249 families.
• Consistent style: one whole bird per image, mostly in side view, same rendering and background throughout; the redrawn images also share one pose.
• Linked identifiers: each species has its scientific name and the IDs used by eBird, Avibase, GBIF, Wikidata and iNaturalist, so you can map it to iNaturalist, CUB or your own label set.
• Quality labels: every image has a review score and a pass/fail flag, so you can filter or weight by quality.

Contents

Field	Description
image	WebP with an alpha channel, about 700 × 1,070 px
id	AviList species ID, e.g. Struthio_molybdophanes
scientific_name, authority	AviList v2026
order, family, genus, seq	AviList v2026 taxonomy and sequence
chinese_name	Simplified Chinese name
english_name, names	AviList English name; names in up to 43 languages from the IOC multilingual list
ebird, avibase, gbif, wikidata, inaturalist	External IDs
form	Empty for the species image; otherwise the subspecies shown
review_score	0–100, from the review described below
meets_quality_bar	true if the image passed the strict review
reference_type, reference_page, reference_author, reference_license	The reference used to draw the image (photo, board of several images, or written description only)

Splits: a single train split with all 11,064 images. Filter on meets_quality_bar if you only want the images that passed review.

How the images were made and reviewed

1. For each species we collected a reference photo or, when no single photo showed the field marks well enough, a board of several images. We also wrote a short description of the field marks: plumage, bill, eye, legs and feet, and proportions.
2. The image model drew the species in a standard pose we set. The pose, background and lighting of the reference photos were not used.
3. A separate model reviewed each image, playing the role of a field-guide plate editor. It scored five aspects from 0 to 4: identification marks, plumage, bill and eye, legs and feet, and proportions. It also answered one question: could this image serve as the species' plate in a field guide? An image meets the quality bar when all five scores are at least 3, the identification score is 4, and the answer is yes.
4. Images that failed were redrawn with the reviewer's corrections, up to several rounds. The repository is still being improved in this way.

In this snapshot, 9,939 of 11,064 images (89.8%) meet the quality bar. Of the 1,125 that do not, 1,099 scored 70–79 and 26 scored below 70.

Limitations

• AI-generated, AI-reviewed. The review was done by AI models, not by ornithologists. Passing the bar means the reviewer found no key error. It does not mean an expert has checked the image. Expect some errors in fine details, such as toe arrangement, feather tracts on the folded wing, iris colour and fine barring, especially in species that are poorly photographed.
• One plumage per species. Usually this is the adult male in breeding plumage. Females, juveniles and other plumages are mostly missing.
• Synthetic textures. The images can carry artefacts typical of image models. If you mix them with photographs in training, consider domain gaps, as Rodríguez et al. describe for real illustrations.
• Reference photos. The reference photos are not included in this dataset; their pages, authors and licenses are listed per image. Many of them carry non-commercial or all-rights-reserved terms. The illustrations are new drawings of the species, not copies of the photos, but if this matters for your use, the reference fields let you filter.
• Taxonomy. Names follow AviList v2026. Older datasets such as CUB-200-2011 or iNaturalist 2021 use older taxonomies, so some species have been split or lumped since then. Map them through the external IDs and check the edge cases.
• Not for identification decisions. Do not use the images alone to decide what species a real bird is, for example in conservation or legal contexts.

Citation

@dataset{huashu_bird_sticker_2026,
  author    = {Huashu},
  title     = {huashu-bird-sticker: AI-drawn field-guide illustrations for 10,973 bird species},
  year      = {2026},
  version   = {2.2.0},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23968399},
  url       = {https://github.com/alchaincyf/huashu-bird-sticker}
}

Please also credit the stickers as the license asks, for example: "Bird illustrations: huashu-bird-sticker, CC BY 4.0".

Contact

Open an issue in the repository or email alchaincyf@gmail.com. If an image is wrong, tell us the species and what is wrong, and we will redraw it.