From Idea to Image, Part Four

About this series

From Idea to Image is a weekly self-study course in making images with AI. The text of the series is also written with the help of AI and is edited and published by Owrang.

Each part teaches one skill and ends with a simple exercise you can do the same day. The images on this page were made with the same description and the same scenes you will read below.

Put last week's four images side by side once more. You gave each one a word for its feeling, and if the exercise went well, the four words were different. Now look at them with another question. Is it the same woman in all four? In ours it was not. For the fourth image we asked twice for the same woman, and twice a different woman sat down at the table. If the woman changed in your images too, you made no mistake. You reached the limit that this week is about.

One image can stand alone. But a story, a children's book, a set of posters for one event or a family album that never existed needs a person who can walk from image to image and stay herself. This week we learn how to help the machine hold on to one face. Our method rests on three things, words, an image and visual markers, and none of them depends on a particular tool.

Why the face changes

When you write "a woman in her sixties", you have not described a person. You have described millions of people. Unless you give the machine a reference, it has to build the face from scratch every time. Every image is made anew from the words you give it, and wherever the words leave a choice open, the machine chooses again. Last week almost everything was open. We gave her age and nothing else.

A novelist has the same problem and solves it the same way. A character who is only "a woman in her sixties" has no face for the reader either. The face appears when the writer gives her a narrow face, a white streak in her hair and glasses she keeps pushing up her nose.

Words: the character description

Write a short description of your person and treat it as fixed. Five to seven visible features are enough, the same number we used for a scene in Part Two. Go from the general to the particular. Age. The shape of the face. The hair, its color and how it is worn. Eyes and eyebrows. One mark that belongs only to this person, a white streak, a mole, a gap between the front teeth. Then the clothes.

Leave out the words that cannot be seen. Kind, dignified and tired are feelings, and as we saw in Part Two, the machine has to guess what they look like, and it guesses differently each time. Leave out the word beautiful as well. It often pulls a face toward the same smooth, unmarked average.

Once the description is written, do not rephrase it. Copy it word for word into every prompt. "Gray hair gathered at the back" and "silver hair tied up" are the same to us, but to the machine they are two separate instructions.

Image: the reference image

Words narrow the choice, but they do not close it. Many faces fit even the most careful description. So before you make any scene, make one plain portrait of your person and call it the reference image. From then on, that is who she is.

A good reference image is an uneventful one. The face is turned to the camera, the light is soft and even, the background is plain, the face is calm, and nothing covers it. Any strong effect in the reference image usually travels with the face into the images that follow.

Many tools let you give an image along with your words, or continue working in the same conversation where the image was made. Wherever one of these is possible, use it, and say plainly what you want from that image. "The same woman as in this image." Where neither is possible, the description has to carry all the weight, and you should expect a sister who resembles her, not the woman herself.

Visual markers: what she wears and carries

Think about how we recognize a character in a comic strip or an old film. Often it is not the face at all. It is the hat, the coat, the mustache, the cut of the hair. A viewer who sees the same round glasses, the same green cardigan and the same white streak accepts small differences in the face and does not even notice them.

So give your person two or three fixed markers and keep them in every image, even where the scene would allow a change. Later, when the face holds well, you can change her clothes. For now, the cardigan is part of her face.

This week, step by step

We take the woman from last week's kitchen, and this time we make her one person. The description stays untouched in every step, and only the scene changes.

Step 1. Write the character description. Ours is this. "A woman in her sixties with a narrow face, high cheekbones and dark eyes under thick eyebrows. Her gray hair is gathered in a low bun, with one white streak at the front. She wears round metal glasses, a dark green cardigan over a cream blouse and a thin gold chain." From here on, every prompt in this lesson carries these three sentences. Wherever we write "the description", put them there word for word.

Step 2. Make the reference image. "A portrait of this woman." The description. "Head and shoulders, at eye level, facing the camera, with a calm face. Soft daylight from a window and a plain light wall behind her. Vertical frame, like a photograph." If you can, make it several times, choose the face closest to the one in your mind, and save it under a name. Ours is called Mahin. The name is for your notebook, so that you always know which image is the reference.

Step 3. Take her to the kitchen. Give the reference image to the tool, or continue in the same conversation, and write. "The same woman as in the reference image, with the same face, the same hair, the same glasses and the same clothes." The description. "She sits at a kitchen table holding a glass of tea, with a window beside her. Medium shot at eye level. Soft daylight from the window on a cloudy morning. A horizontal frame, wider than it is tall." If your tool does not accept an image, drop the first sentence and begin with the description. This is the first image. We start with a medium shot on purpose. A face that takes up more of the frame usually keeps its features better than a small one.

Step 4. Compare before you go on. Put the first image beside the reference image and look at five things. The shape of the face. The hair and its white streak. The glasses. The clothes. The age. If one of them has slipped, correct only that one. "Keep everything as it is. Her hair is gray with one white streak at the front, gathered in a low bun, as in the reference image." If the description does not hold in this simple scene, it will not hold in the harder steps.

Step 5. Change only the light. This is the image that got away from us last week. Go back to the reference image, not to the image you have just made, and write. "The same woman as in the reference image, with the same face, the same hair, the same glasses and the same clothes." The description. "It is night. She sits at a kitchen table holding a glass of tea. The only light is a single bare bulb hanging above the table, hard light with deep shadows, and the window is dark blue. Medium shot at eye level. A horizontal frame, wider than it is tall." This is the second image. Hard light from above is a real test, because shadows redraw a face. If she stays herself here, your description is working. In our image she stayed herself under that bulb. The kitchen may not be the same kitchen as in the first image. This week that does not matter. One thing at a time, and this week the thing is her face.

Step 6. Change the place and the distance. "The same woman as in the reference image, with the same face, the same hair, the same glasses and the same clothes." The description. "In the morning she stands in a small courtyard, watering pots of geraniums with a metal watering can. Wide shot at eye level, with her whole figure in the frame. Soft morning light. A horizontal frame, wider than it is tall." This is the third image. In a wide shot the face is small and the machine has little room for it. Here the markers do the work, and if they are not enough, move one step closer. The green cardigan, the bun and the glasses tell us who she is before we can read her face. In our image this step was not completed. We asked for a wide shot, and the tool stood closer and cut the frame above the knee. The face held, but the test of distance was not carried out.

Step 7. Put the reference image and the three images side by side, without the prompts, and show them to someone who has not followed your work. Ask how many people are in these four pictures.

When the face goes its own way

Does she look ten years younger? Many tools lean toward younger, smoother faces. Name the signs of age as things that can be seen, the lines around the eyes and mouth, the skin of the neck and hands, and take out any word like beautiful or elegant. The woman in our images also came out a few years younger than sixty.

Does each image drift a little further from the reference? You are probably building each image on the one before it, and small differences add up, like a message whispered from one person to the next down a line. Go back to the reference image every time.

Did the face change when the angle changed? A face from the side or from above is a harder question for the machine than a face from the front. Change the angle a little at a time, and if your tool accepts more than one image, make a second reference image, the same woman in three-quarter view.

Did the background or the light of the reference image come along with her? Say what you are taking from that image and what you are not. "Take only her face, hair and clothes from the reference image. The place and the light are new."

A real face

Everything in this lesson is about a person who does not exist. If you want to use a photograph of a real person, yourself, your mother, a friend, the method is the same but the question is different. An image of a real person in a place they have never been, doing something they have never done, can be unwelcome to them or give a false picture of them, even when it is made with affection. Ask first. Use your own photographs or ones you have permission to use, and read what your tool says about uploading pictures of people, because the rules differ from one tool to the next. And when you show such an image to others, say that a machine made it.

Working vocabulary

character description, the fixed written portrait of a person

reference image, the portrait every later image is compared with

head and shoulders, a portrait that ends below the shoulders

facing the camera, the face turned straight toward us

three-quarter view, the face turned partly to one side

profile, the face seen from the side

This week's exercise

Choose one person, an invented one, not a real one. Write the description in five to seven visible features. Make the reference image and save it under a name. Then make three images of the same person. One in a room in daylight, one in the same room at night, and one outdoors in a wide shot. Go back to the reference image each time, and do not touch the description. Keep the description, the reference image and the three images together.

Check your work

Show the four images to someone without the prompts. Do they count one person? Can they name two things that told them so? Now cover the faces with a finger. Can you still tell it is her? If they count two people, find the image that holds the stranger and compare its prompt with the others, word by word. The difference is usually there.

The aim this week is not a perfect twin. It is a person the viewer recognizes.

Next week. The medium of the image. Photograph, watercolor, gouache, pencil drawing, print. What happens to the same scene when its medium changes?

The images on this page. In the middle of the page, the reference image and the kitchen in daylight. At the end, the kitchen at night under a single bulb and the courtyard. To fit the page, each pair was made as one two-panel frame. The face, the hair, the white streak, the cardigan and the chain held in all four panels. The frame of the glasses is darker in the daylight kitchen than in the reference image, and the courtyard, as we said, did not become a wide shot.