How to Make a Large AI Illustration Set That Holds Up: The 11,064 Bird Stickers Behind a Five-Point Review Checklist
Somebody just drew every living bird species on Earth, all 11,064 of them, and published a list of the 3,394 drawings that are not good enough.
That second number is the remarkable part. Big AI image sets usually arrive as a gallery of best results, with the failures deleted and the hit rate never mentioned. The huashu-bird-sticker project ships every sticker with a score and a plain pass-or-fail flag. Seven in ten passed. The rest are labelled.
If you make illustration sets for a living, the stickers are nice. The scoring system is the thing worth stealing.
What the project is
The stickers were drawn for Mingji, a bird-watching notebook app for iOS, and put on GitHub, the public code-sharing site, on October 30 by the maker, Huashu. There is one sticker per species on version 15.2 of the IOC World Bird List, a standard scientific roster of living birds. Each one is a transparent image with a white outline, about 1,000 pixels on its long side, filed by order, family and species. A data file gives each bird's name in 43 languages.
The project's front-page notes (the README) describe the production in one breath: for each species, "we found a reference photo," then "drew the sticker with GPT-image, upscaled it, and cut it out." GPT-image is OpenAI's image model; upscaling enlarges the picture, and cutting out removes the background. Then the important sentence: "Every AI sticker was then scored by a separate reviewer against the reference for field marks."
Field marks are what birders use to tell species apart. The project scores five of them: the bill, the eye, the legs and feet, the pattern of the feathers, and their colour.
The position: generation is the easy half
An image model will draw you a bird in seconds. It will draw it confidently, beautifully and slightly wrong. The bill comes out a touch too long, the eye ring disappears, the wing bars change colour. To a casual viewer it looks fine. To anyone who knows the subject, it is a different bird.
This is the problem every illustrator hits with AI sets, whatever the subject. A product line where the logo drifts. A character roster where costumes lose their details. A set of regional food icons where the dishes stop being recognisable to people who eat them.
The bird project solves it by refusing to judge the drawings by feel. "Does this look right?" is a question you answer differently at 9 a.m. and at 11 p.m., and differently on drawing 40 and drawing 4,000. "Is the bill the right shape compared to this photo?" is a question you can answer the same way every time.
That is the whole trick. The checklist turns taste into a repeatable test, and the repeatable test is what lets the set grow without quality sliding.
How the review works, step by step
Here is the structure as the README lays it out.
A reference for every subject. Each drawing is judged against a real photo, recorded in the data so you can trace it. The example record links to Wikimedia Commons. For 453 species, the README says "no single photo was good enough," so the team used a board of several images plus a written description.
A score against a fixed list. Each sticker gets a number based on those five field marks. The example in the data scores 91.
A pass mark. Eighty or above passes. Below 80 gets redrawn.
A cap on redraws. The README says failing stickers "were redrawn up to the limit, and we kept the best-scoring version." Every species still gets a picture, which matters for an app that has to show something for every bird.
A blind second look at the borderline. Stickers scoring 80 to 82, right on the line, went through a second review without the first score visible. Of 852 borderline stickers, 211 failed that second look and were redrawn.
An honest label on the result. Every sticker carries its score and a true-or-false flag that says whether it met the bar. The final count: 7,670 pass, 3,394 do not, most of those scoring in the 70s.
The blind second look is the step most people would skip, and the numbers show why it matters. A quarter of the drawings that scraped past the first review did not survive a fresh pair of eyes. If you only check things once, the borderline is where your errors live.
Put this into practice on a set of fifty
You do not need eleven thousand subjects to use this. It works at the scale of a client job: fifty product icons, thirty character cards, twenty-four illustrations for a children's book where the same fox has to stay the same fox.
1. Write the checklist before the first prompt. Pick four to six features that make each subject recognisable. For a character set (working from a model sheet, the page that shows a character from every angle) that might be face shape, hair, the signature costume piece, colour palette and the prop they always carry. For products: silhouette, logo placement, the one detail the client always mentions. Keep it short enough to check in under a minute per image.
2. Collect one reference per subject. A photo, a model sheet, the client's own product shot. Note where each came from. If a single image does not show everything, do what the bird project did and make a small board with a written note.
3. Score each feature, then total it. A simple version: give each feature 0, 1 or 2 points (wrong, close, right). Five features gives you a score out of ten. Decide the pass mark in advance. Writing the number down before you see any images keeps you honest when you are tired and want to be done.
4. Redraw the failures, with a limit. Three attempts is a reasonable cap. Keep the best one even if it never passes, and mark it. A limit stops you from spending an evening on one stubborn subject while forty others wait.
5. Re-check the borderline blind. Take every image that passed by a single point, shuffle them, hide the old scores, and score them again the next morning. Or hand that batch to a colleague with the checklist and nothing else.
6. Keep the scores in a spreadsheet. One row per subject: file name, reference, score, pass or fail, number of attempts. It takes minutes. It also becomes the most useful thing you can show a client, because it answers "how do you know these are right?" with a table instead of a shrug.
7. Ship the failures labelled, or not at all. If a client needs every subject, deliver the best attempt and flag it. If they do not, leave it out. What you never do is let a failed image sit in the folder looking identical to a passed one.
Where it breaks
The README leaves out the tools. It does not name the reviewer, the upscaler or the cutout tool, and it gives no cost or time. The README never says whether a person or an AI model did the scoring. That matters. If the same kind of model that drew the image also judges it, it can share the same blind spots. If you copy this workflow, decide who reviews, and be aware that a reviewer who is too similar to the artist catches less.
A score is only as good as the checklist. The five field marks catch anatomy and plumage errors. They do not catch posture that looks wrong, a style that drifts across the set, or a drawing that is accurate and lifeless. Your checklist will have gaps too. Look at the whole set side by side at the end, and add a feature to the list when you notice a pattern of misses.
References carry their own rights. The bird project records its reference sources but does not ship the photos. If you build a set from reference images, especially a client's or a stock library's, check whether you are allowed to use them to guide AI drawings.
Accuracy has a ceiling. The README says it plainly: "These are illustrations, not field-guide plates." It tells readers to check a field guide before relying on a fine detail. Even passing stickers are friendly pictures, not scientific records. Say the same about your set, out loud, to whoever commissions it.
The license has a fence. The stickers use PolyForm Shield 1.0.0. In plain terms, you can use, edit and redistribute them, including in paid work, as long as you credit the project with a link to its GitHub page and keep its copyright notice. The data files are separately MIT licensed. The one exception is any birding, bird-identification or field-guide app, site or service. Put a heron on a poster or a café menu and you are fine. Build a bird-ID app with them and you are not.
It is heavy. The full download is about 700 MB. If you just want a few birds, the README suggests loading single stickers by URL or downloading only the folders you need.
What to take from it
Open the repository and pick any sticker with a score in the low 70s. Then find one in the 90s from the same family and compare them. You will see exactly what the checklist caught, and it will probably be smaller than you expect: a bill a little thick, an eye a little high.
That is the level of detail your clients notice and AI gets wrong without warning. The cure is not a better prompt. It is a short list, written down before you start, and the discipline to score against it every single time.
- Sources: huashu-bird-sticker on GitHub · README · PolyForm Shield 1.0.0