How to take great reference photos for AI illustrations

If you only do one thing carefully in the StoryInk flow, do this. The reference photos are the single largest variable in whether the illustrations look like the person or pet you are making the book about. Better photos in, better book out. The prompt matters too, the mode matters too, but neither of them can compensate for blurry low-light selfies as the reference set.

We have looked at a lot of AI illustrations by now — books about toddlers, about grandparents, about a partner, about the family dog — and the patterns are pretty consistent across all of them. This post is the short version of what actually moves the needle, whoever your subject is.

Natural light: the biggest factor in any reference photo

The single biggest factor. Outdoor daylight on an overcast day is ideal — even, soft, no harsh shadows. Sunny but shaded is the next best. Direct midday sun is usable but it blows out highlights and flattens skin tone or fur color, which can confuse the AI about what your subject actually looks like.

Indoor photos taken next to a large window are fine. Indoor photos taken with overhead lights or, worst, with your phone's flash, are not. The flash flattens the face and reflects in the eyes, and the AI tends to render the result in an oddly waxy or metallic way. Skip flash entirely.

Eye level

Phone-held-above-and-angled-down is the default snapshot pose. It is also the worst angle for a reference photo. The head looks oversized, the body foreshortens, the perspective is wrong. The AI absorbs that distortion and bakes it in. This is just as true for a photo taken looking down at a small child or a dog as it is for a selfie held overhead.

Get to your subject's eye level. Crouch down to the toddler. Sit across the table from Grandpa. Lie on the floor for the dog. The reference photo should look like it was taken by someone standing or sitting right in front of them. That is the perspective that produces an illustration that looks like them.

Multiple angles

One angle per upload, ideally three to five photos total covering different angles. The combinations that work well:

• One clean front-on portrait, head and shoulders, eyes visible.
• One full body in profile (side view, standing or sitting).
• One three-quarter angle of the face.
• One that clearly shows any distinctive feature — a hairstyle, glasses, a coat pattern, a favorite outfit.
• One in an active pose if you have it (walking, laughing, mid-play).

You do not need all five. Three good ones beat five mediocre ones. The system uses the references collectively to build up a picture of what your subject looks like from any angle — and inside a book, the system uses page 1 as a character anchor for the pages that follow. An angle the references do not show is an angle the AI will invent.

Bursts are your friend

People and pets move. The single click of the shutter at the moment your toddler finally looks at the camera is the moment they also blinked. Use burst mode (hold the shutter button down on iPhone / Android). You will get thirty frames in five seconds, two of which will be sharp and one of which will be perfect.

This sounds like a small thing. It is not. Most of the bad reference sets we see could have been good if the photographer had taken five times as many photos and picked the best of each angle.

Pose variety beats pose quality

Three photos of your kid sitting on the couch are less useful than a sitting-on-the-couch photo plus a standing-in-the-yard photo plus a curled-up-reading photo, even if the couch photos are individually a little prettier. The AI is going to need to draw your subject in poses you did not photograph, and it builds up its ability to do that by seeing them in different ones.

What confuses AI illustrations

Direct list of things to avoid in reference photos, based on the failure modes we see most often:

• Sunglasses, masks, or hats that hide the face. The AI sometimes reads the accessory as part of the person and tries to draw it on every page.
• Full costumes — Halloween, holiday, a dog in a sweater. Same problem. Show your subject as they actually look for the references.
• Extreme close-ups where only one eye or only the nose is in frame. The system needs to see the whole face structure.
• Photos with another person or pet in the frame. The AI sometimes cannot tell which one you mean. One subject per reference photo.
• Heavy filters, beauty modes, or Snapchat-style edits. The face structure gets distorted.
• Black-and-white photos. The model needs hair, skin, and clothing or coat color.
• Very small or very pixelated photos. If you can barely see your subject clearly, the AI definitely cannot.

What helps AI illustrations look like your subject

And the inverse list. Things that consistently produce better illustrations:

• Your subject against a relatively plain, contrasting background. A person on a clear porch is easier to parse than a person in a busy crowd.
• Sharp focus on the eyes. Eyes are where the AI most strongly anchors character identity.
• Calm poses where the head and body are both visible and in plane. Action shots are great too, but they should be in addition to, not instead of.
• Recent photos. If you are making a book about Grandma now, recent photos are more useful than photos from twenty years ago. The AI will lean on whatever it sees.
• If there is a strong distinguishing feature — a particular smile, freckles, a beard, a coat marking — a reference photo that shows it clearly is worth its weight in gold.

The five-minute version

If you have five minutes with your subject and want to take the best possible reference set: find a quiet spot with soft, even light. Get to their eye level. Burst-shoot a portrait of the face. Burst-shoot a profile of the whole body. Burst-shoot a three-quarter angle. Pick the sharpest from each burst. You have three excellent references. That is plenty.

Five minutes spent here will improve your book more than any other five minutes you could spend in the flow. Worth doing before you start.