How AI illustrations are made

An image model knows what things look like, not what names mean. Ask one for a species it doesn't know and it paints something plausible and wrong. So Featherframe never leaves the image model to guess. Each illustration is made in five steps.

1. The AI ornithologist

Before anything is drawn, a text model plays a naturalist briefing an illustrator who has never seen the species and can't look it up. In at most 80 words, it says:

• What group the species belongs to, in everyday terms.
• Its body plan and true size.
• The features that tell it apart from similar species.
• How it carries itself.

It also lists four to six real plants tied to the species: food, host, nest site, or habitat. Each comes with its Latin name, a season, and how it looks at that one stage of its year: leaf shape, growth habit, and whether it carries bud, flower, or fruit. The list leans toward flowering shrubs, vines, brambles, and herbs where that's true, the way Audubon's plates, the sheets of his book, do, instead of filling up with trees.

Last, it says whether the species is a bird. BirdNET also hears katydids, cicadas, and bats. Before the brief, they came back with feathers.

The ornithologist is asked to state only established fact. Its brief is written once per species and kept.

2. Art direction

Each illustration draws its own art direction, like rolling dice with weighted sides:

• Subject: most often a plain portrait of the species. Sometimes it's feeding, at its nest with eggs or young, beside a juvenile, or in a dispute. Audubon painted all of these. Nests and juveniles are for birds only.
• Figures: one, two, or three. A second figure has to show something the first can't, like another side of the species or another motion.
• Support: a long diagonal branch, an S-curve, a single upright stem, or open air.
• Plant: one from the ornithologist's list. A nest or a juvenile gets a plant in its spring or summer stage, so nestlings don't sit in autumn leaves.
• Mood: how full the plant is, how still or dramatic the moment is, whether moss or lichen appears, and whether the figures notice each other.

The odds favor restraint: most illustrations are fresh, spare, and calm. The direction is drawn at random because a fixed instruction is obeyed every time. An early prompt asked for a few natural imperfections, and nearly every leaf came back tattered.

Regenerate rules out the last illustration's subject, figure count, fullness, support, and plant, so a new illustration looks new.

3. The style prompt

The rest of the prompt describes how Audubon's plates were made, not how they look at a glance. It was written from a study of 120 of the plates, then revised after rounds of blind tests against real plates. Each line removes something that gave a fake away. It asks for:

• The process: watercolor washes over engraved lines and aquatint grain, gently uneven, never airbrushed. Nothing is drawn finer than a brush and an engraving tool can hold. Eyes are small and matte, and never look straight at you.
• Color: drab bodies with one or two confident accents, muted sage greens, whites left as bare paper, and one small paintbox for the whole sheet.
• Composition: figures on one shared branch at staggered heights, with half to two-thirds of the sheet left bare. The species is always shown plainly, with the features that identify it. It's never hidden or camouflaged.
• Setting: one real plant drawn to a botanist's standard. Woodpeckers get dead wood, ground species a strip of ground, waders a low shore, and fliers open air. The whole sheet is one moment in one season.
• Anatomy: true feet, claws, and wing feathers for birds, and the same care for any other animal: its legs, segments, and wing veins.
• No text except small figure numbers.

4. Audubon's own plates as references

The image model also gets three of Audubon's illustrations, at reduced size. Similar species come first: a duck gets waterfowl, an owl gets owls, a hawk gets hawks. Otherwise, it gets airy sparrow plates. In testing, busy plates like the Cardinal pulled illustrations toward heavy foliage. The references are shuffled, so regenerating doesn't copy the same sheets.

5. Paper and caption

Featherframe matches the paper color to the screen. It doesn't crop the illustration.

The caption is set like Audubon's: the title in script, the Latin name, and a legend that names the figures and the plant, like "Male, 1. Female, 2." and "Chestnut Oak. Quercus prinus." Where Audubon's illustrations show their number, like "Plate CLIX", AI illustrations show ✦. The other corner names the artist, like "John James Audubon". On an AI illustration it says "AI illustration after Audubon".

Featherframe saves each illustration's brief, art direction, plant, and references with it, and they're included in backups.

If the image can't be generated, the species gets its name card.

AI collages

An AI collage follows Audubon's late composite plates, where several species share one sheet. The prompt lists the day's species, most-heard first, each with the first line of its brief. Every species gets its own place on one branch, set in the season of the collage's date, or in the day's weather (Tree branch under Collage; Bare keeps it bare), drawn to true relative scale: a crow dwarfs a wren. The most-heard species gets the most prominent place, and each figure is numbered to match the key under the collage. With seven or more species, the branch divides to fill the whole sheet.

Its references are three of Audubon's own composite plates, such as five warblers up one sparse spray.

For costs, providers, and privacy, see AI illustrations.