Работа с референсами

A reference isn't just for aesthetic purposes. It takes the work that the AI can do on its own out of the prompt. And it does this better than a long list of words.

One photo—is that enough?

Sometimes one photo is enough. If you’re creating a portrait based on another portrait—a face, shoulders, or part of the figure—that’s enough. But a single photo shows only one side. The AI will fill in what’s outside the frame. It might guess correctly. It might not. Usually, it’s the second option.

If you don’t have any other photos, describe the figure in character's binding prompt. Height, proportions, build, distinctive features. Be detailed. Text won’t replace an image, but it will give the model at least some guidelines. Sometimes that’s enough to keep the result from veering into random anatomy.

The optimal set

For stable results, you need at least three headshots—front-on, to the left, and to the right—and at least one full-body photo in a “swimsuit” that shows almost your entire figure. Ideally, 3–5 headshots and 2–3 full-body photos. That way, there are almost no blind spots left for the AI.

It’s not about quantity, but about variety. Four identical selfies are useless. One full-face shot, one in profile, one from a different angle, and one full-body shot—that’s already a solid working base. Each photo should add new information, rather than repeat the previous one.

You can submit photos separately. In the bindig prompt, specify which photo to use for each part. Face—from the first, profile—from the second, body—from the third. This reduces confusion. That way, the AI doesn’t have to figure out on its own whose nose belongs where.

For frequently used characters, it’s convenient to gather capsules and cards. A capsule is a single image showing the character’s basic appearance—specifically, three faces and one body. A character card is an expanded set where a single image shows the character not only from different angles but also in different poses. Once you’ve gathered them, the character’s appearance remains virtually consistent. After that, you can change the scenes, lighting, and clothing, and the character remains true to themselves.

A reference doesn’t decorate the prompt—it economizes it. If you need a butt, but the reference only shows a full-face view, the model won’t guess correctly. It’ll draw something—not necessarily what you wanted. So it’s better to spend ten minutes on a proper set of references than to spend time later sifting through generated images and dealing with random anatomy and banned content.