There is a persistent frustration that haunts anyone who spends time generating AI portraits. The faces are often stunningly beautiful. The skin is flawless. The lighting is cinematic. And yet, the expression feels hollow. It is a smile that does not seem to come from anywhere. It is a sad face that looks more like a static mask than a moment of genuine feeling.
After testing between four and five hundred images to observe how different prompts actually affect the final output, one thing became very clear: simply writing "smile" or "sad" into a prompt is not enough. The AI delivers the mechanical movement of muscles, but it rarely delivers the emotional truth behind them. This article is the result of that testing. It is a practical guide to moving beyond flat expressions by changing how we think about prompts entirely.
Why Modular Prompts Fail
The first instinct for many users is to treat the face like a control panel. If you want a character to look sad, you describe the eyebrows. If you want a softer gaze, you describe the eyelids. If you want a smile, you break it down into closed lips, open lips, or the specific angle of the mouth. This modular approach seems logical because it gives the AI explicit instructions for each part of the face.
In practice, the results are often unsettling. While the AI follows each individual instruction, the combined effect can look stiff, unnatural, or even eerie. A character might have perfectly raised eyebrows and a correctly shaped mouth, but the overall expression feels like a patchwork of unrelated parts rather than a unified human emotion.
There is a well-known observation about the beauty of certain actresses: if you take each of their facial features individually, none of them are perfect. But when those features are combined on one face, the result is extraordinary. The modular approach to prompts works in the opposite direction. You write what you think are perfect descriptions for each part, but when the AI assembles them, the whole is far less than the sum of its parts.
This method also creates a serious practical problem. Writing modular prompts is difficult for a human to do well. Describing exactly how each muscle moves during a complex emotion requires a level of anatomical precision that most people do not possess. And when the output is wrong, debugging the prompt is nearly impossible. You cannot easily tell which part of your description caused the problem.
A Better Approach: Narrative-Driven Prompts
The alternative that emerged from testing is fundamentally different. Instead of describing the physical appearance of the face, you describe the reason for the emotion. This means writing a short story that explains what happened to the character, how she feels about it, and who she is feeling that emotion toward.
Early experiments included using emojis or kaomoji (text-based emoticons) to help guide the expression. These were rejected because they introduced a weight interference. The symbols competed with the narrative for the AI's attention, and the resulting expressions lost the unique emotional quality that the story was trying to create.
The final structure is simple and uses only natural language:
What happened + How she feels + Who the feeling is directed at.
Consider a concrete example. Instead of writing "sad, disappointed, looking down," you write something like this:
She prepared a gift carefully, but the person she cares about most put it aside without a second glance. She looks at the person behind the camera, feeling hurt and disappointed, yet still hoping for a response.
The resulting image does not literally show the gift or the neglect. But the expression carries a specific weight. The eyebrows and mouth show a low, disappointed quality that feels real because it has a cause. The character is not just "sad" in a generic sense. She is sad in a very particular way that comes from being ignored by someone important.
Basic Template for Single Emotions
For emotions that are relatively straightforward, a basic template works well:
[Character] just [event]. [Character] looks at [target], feeling [primary emotion], and [secondary emotion or expectation, optional].
Here is a tested example:
She finally sees someone she has missed for a long time. She looks at the person behind the camera, feeling joy and relief, as if a long wait has finally been answered.
The key is that the event should be small and recent. It should feel like something that just happened a moment ago. "She finally sees someone she missed" works because it is a specific, immediate trigger. A larger, more abstract event like "she has had a difficult life" is too broad to give the expression a clear anchor.
Handling Complex, Mixed Emotions
The real challenge comes when a character needs to feel two things at once. A person can be happy and sad simultaneously. They can be angry but also deeply hurt. They can smile while their eyes are full of tears. Writing "happy, sad, crying, smiling" into a prompt does not work because it does not tell the AI how these emotions relate to each other.
To handle this, you must specify two things. First, you need a hierarchy: which emotion is dominant? Second, you need a motivation: why is the character showing one emotion while feeling another?
Consider a reunion that is also a farewell. The character finally meets someone she has been longing to see, but she also knows this will be their last meeting. The joy of reunion and the grief of farewell exist at the same time. If you only write "happy and sad," the AI might give you a neutral expression that captures neither feeling well.
Compound Template for Inner vs. Outer Emotions
For these complex situations, a compound template provides the necessary structure:
[Character] just [event]. Inside, [character] feels [primary emotion], but also [secondary emotion]. Because [reason for hiding or expressing], [character] looks at [target] and tries to [external behavior]. Yet [one or two visible details] reveal [hidden emotion]. It is a look of [summary of emotional conflict].
Here is a tested example that uses the reunion-farewell scenario:
She finally meets the person she has been longing to see, only to learn that this meeting is also their final goodbye. The reunion brings a flicker of joy, but inside she feels more sadness and reluctance. Because she wants the other person to remember her as happy, she looks at the person behind the camera and tries to smile. Yet tears roll down her cheeks, and her lips barely hold the smile. It is a look of sadness wrapped in a brave attempt not to worry the other person.
In the resulting image, the smile and the tears coexist. The viewer can see that the character is trying to be happy, but the sadness is winning. The prompt works because it explains the motivation behind the conflicting expressions.
Another example shows anger mixed with hurt:
She finally opens up to the person she trusts most, only to be dismissed with a casual "you are overthinking it." She feels anger, but more than that, she feels hurt that her true feelings were not understood. Because she does not want to appear weak, she looks at the person behind the camera and tries to stay strong and unyielding. Yet her brow is tight, her lips tremble slightly, and her eyes begin to redden. It is a look of wanting to confront the other person, but being afraid that she will break down if she speaks.
In this image, the visible details are minimal: a tight brow, trembling lips, reddening eyes. But those few details are enough to communicate the internal conflict between anger and hurt.
Practical Tips and Templates for Readers
The core lesson from all this testing is simple: to get rich, natural expressions, you need to think about the story behind the emotion. Do not ask yourself "what should her face look like?" Ask yourself "why is she feeling this way?" and "what does she want the person looking at her to see?"
Here are the two templates ready for use. Replace only the expression part of your existing prompts. Keep your character descriptions, clothing, scene, and style settings unchanged.
Basic Template (for single or naturally paired emotions):
[Character] just [event]. [Character] looks at [target], feeling [primary emotion], and [secondary emotion or expectation, optional].
Compound Template (for inner vs. outer emotions):
[Character] just [event]. Inside, [character] feels [primary emotion], but also [secondary emotion]. Because [reason for hiding or expressing], [character] looks at [target] and tries to [external behavior]. Yet [one or two visible details] reveal [hidden emotion]. It is a look of [summary of emotional conflict].
One important warning: the event description can inadvertently change the scene. If you write "she just received bad news," the AI might add a phone or a letter to the image. If you want to change only the expression, you can add a constraint such as "the event is only for explaining the expression; keep the original character, scene, and composition unchanged." Then check the output to see if the instruction was followed.
You do not need to fill every blank in the compound template every time. If there is no hiding or masking, do not force it. If two emotions flow naturally together without conflict, the basic template is sufficient. Choose the visible details that are most relevant to the emotion you want to convey. One or two details are usually enough.
Conclusion: The Power of Story in Prompts
The difference between a flat expression and a compelling one is not about finding better synonyms for "smile." It is about giving the expression a reason to exist. When you write a prompt, think like a storyteller. What just happened to this character? What is she feeling right now? What does she want the person in front of her to see?
The next time you want to generate a character who is "crying while smiling," do not just list the two emotions. Ask yourself: why is she smiling when she is sad? What is she trying to hide? Who is she trying to protect? The answer to that question is what will make the expression real.
Try the templates in your next generation. Replace the expression part of your prompt with a short narrative. See how the face changes when it has a story to tell. Share your results and let the community learn from what works. The most powerful tool in prompt engineering is not a longer list of adjectives. It is a single, well-chosen reason for the emotion.
