Generating illustrations

The collage art is generated, not hand-drawn. The repo ships 498 kachō-e illustrations (249 species, a perched and a flight pose each). To restyle them or build a set for your own region, the pipeline is four scripts in this directory.

Pipeline

pregen.py renders each bird with Gemini 2.5 Flash Image, on a flat cream ground.
cutout.py removes the ground with BiRefNet and crops to the bird.
build_masks.py rebuilds the collage silhouette masks inlined in apt.js.
verify.py (optional) runs an adversarial species-ID + anatomy check.

pip install -r requirements.txt
export GEMINI_API_KEY='your-key'

# 1. generate (cream ground) for your region's species
python3 pregen.py --labels ~/BirdNET-Pi/model/labels.txt --ebird-region US-CA

# 2. cut the ground off and crop
python3 cutout.py

# 3. rebuild the collage masks, then bump SKETCH_VERSION + IMG_VERSION in apt.js
python3 build_masks.py

--labels takes any Sci|Com per-line file (BirdNET-Pi's labels.txt works directly). --ebird-region filters to species actually seen in your region (needs EBIRD_API_KEY). Re-render one bird with --species "Calypte anna|Anna's Hummingbird" --force.

Why a cream ground

The image model can't cut a clean transparent background on its own: it leaves holes and fringes, worst on pale birds. Rendering on a flat, consistent cream ground gives a known color that BiRefNet removes cleanly, and the steady ground also holds the painting style together across the whole set. cutout.py is the step that makes the backgrounds transparent.

The prompt

prompt.template.md is the kachō-e prompt, sent verbatim per request with {sci_name}, {com_name}, and {pose} substituted. Edit it to change the style. pregen.py attaches up to three reference images per request:

Anatomy (IMAGE 1): a Wikipedia photo of the target species, auto-fetched and cached in assets/references/. Anchors identity and markings. Drop your own references/<slug>.jpg to override.
Anti-reference (IMAGE 2, optional): a photo of a look-alike the model drifts toward, captioned with what NOT to copy. Wired for blue corvids (vs Blue Jay) and swallows (vs Barn Swallow); add more in the ANTI_REFS table and place photos at references/_anti_<key>.jpg.
Style (IMAGE 3, optional): a real Edo-period kachō-e print whose painting technique is borrowed. The genus-to-print mapping is in pregen.py's STYLE_REFS. The prints are not bundled (they are someone else's art); put your own in assets/references/styles/. The Koson and Yoshida prints used originally are easy to find on the public web by the filenames in STYLE_REFS.

All three degrade gracefully: a missing reference is simply not attached.

Hard species

species-notes.json holds one-line diagnostic addenda for species the model gets wrong. Each note names the field marks that matter and the look-alikes to avoid, and is appended to the prompt for that species. Add entries as you find drift; they carry forward to every future regeneration of that bird.

Verifying

verify.py sends each illustration back through Gemini Vision without telling it the target species, then checks the guess, the wing/leg/tail counts, and whether a stray perch crept in. It catches drift a quick eyeball misses.

python3 verify.py --labels labels.txt              # whole library -> verify-results.csv
python3 verify.py --labels labels.txt calypte-anna

What actually goes wrong

Sticks. Perched raptors often come back gripping a twig the prompt forbade. Generate 2-3 and keep the clean one.
Species drift. The model collapses an uncommon species toward a common look-alike (a swift becomes a swallow). Fixes, in order: a sharper species-notes.json note with anti-feature language; an anti-reference; a different style print; a one-off --species regen.
Matched pair. The perched and flight poses must read as the same individual. Review them side by side before locking.

4.0 KiB Raw Blame History