How To Write AI Image Prompts: Tips & Examples

Sometimes you have a clear vision for an image, but the AI model returns something generic or very far from what you have in mind. This is a common challenge in AI image generation, forcing many to guess at “magic words” or give up on their idea entirely.

That’s why we’ve put together this comprehensive prompting guide. We’ve drawn from our own research and testing to equip you with a proven vocabulary and a set of professional workflows, building your skills from the ground up so you can get predictable, high-quality results more often.

We’ll start with the essential building blocks every good prompt needs. Then, we’ll move on to practical, advanced techniques, like how to get more realistic-looking portraits, add text to your images, and use specific camera and lighting terms.

Prompting Essentials: How to Write AI Prompts for Images

Every strong image prompt includes a few core elements. While prompts don’t need to be long or complex, most good prompts are descriptive and clear. A good starting point is to think of three key components: the subject, its context, and the style.

The Bare Minimum: Subject and Context

To get a predictable result, your prompt needs to provide the model with, at a minimum, a clear subject and its context:

• The subject is the “who” or “what” of your image—the main person, animal, or object you want the AI to focus on.
• The context is the “where,” “when,” or “how.” This can be the background (a studio, outdoors), the time of day (golden hour, night), the objects surrounding the subject, or the action the subject is performing.

If you provide just a subject, like “an armchair,” the model is forced to guess the rest. To get a good result, you must provide this descriptive context. For instance, a much stronger prompt would be:

A modern armchair made of light oak and cream-colored fabric, in a bright, minimalist living room.

This prompt works well because it clearly defines the subject (a modern armchair with specific materials) and the context (a bright, minimalist living room).

Image Style

Style is the overall visual aesthetic of your image. This can be a general category (like a photograph or sketch) or something very specific (like pastel painting, 3D render, or isometric 3D).

By default, an AI model will resort to a certain style it has seen most often in its training data for your subject. If you don’t specify a style, you are letting the model choose for you. In our case above, the model chose by itself a photorealistic look, but we can take control by clearly naming a different style:

A watercolor illustration of a modern armchair made of light oak and cream-colored fabric, in a bright, minimalist living room.

There are so many other styles you can use, and that’s why we created a separate comprehensive guide on AI art styles.

Style is a powerful modifier that can dramatically change the mood and meaning of an image. For instance, here is how our armchair image looks in several different styles:

Aspect Ratio

Aspect ratio is the proportional relationship between an image’s width and height. This choice is important for matching your image to its intended use, whether it’s an Instagram post (1:1), a video thumbnail (16:9), a phone wallpaper (9:16), etc.

On Leonardo.Ai you don’t need to specify the aspect ratio in your prompt. You can simply select your desired dimensions from the settings menu before you generate. While you have a few common ratios for quick selection, you can open an advanced customization menu by clicking the Custom button:

Some platforms require you to specify the aspect ratio in the prompt, so always check your platform’s interface first, as this is usually easier and more reliable than trying to force the dimensions with a text command.

Bad Prompts Examples

A common error is to approach AI image models conversationally. Unlike chatbots designed for dialogue, these models are pattern-matching systems. They interpret prompts as a series of descriptive commands, not a conversation.

Conversational filler like “Please create an image of…” or “I would like to see…” is not only unnecessary but can be counter-productive. This phrasing adds linguistic “noise” that dilutes the weight of your critical descriptive terms.

Beyond conversational language, other common errors include being too vague, indecisive, or overly complex:

Prompting for Images With Humans

Generating realistic and compelling images of people is one of the most common goals for creators, but it’s also one of the most complex. Models can struggle with anatomy (like hands) and often produce faces that look strange or artificial.

To get professional-grade results, you need to move beyond simple descriptions and use the language of photography and art direction. This section covers two key areas: controlling the camera’s framing and overcoming the “plastic face” problem for more realistic portraits.

Basic Camera Framings

To control how your subject is presented, you need to specify the camera framing. This tells the AI how much of the subject and their environment should be visible. If you leave this out, the model will guess.

Here are the most common framings you can use:

• Wide shot: Shows the entire subject from head to toe, often including a significant portion of their environment. This is great for establishing a scene.
• Medium shot: Typically frames the subject from the waist or knees up. This is one of the most common shots, balancing the subject with their surroundings.
• Close-up: Focuses on a specific part of the subject, usually the face. This is used to capture detail and emotion.
• Extreme close-up: An even tighter shot that isolates a single feature, such as the eyes or mouth, for maximum dramatic effect.

For example, a text prompt like “A close-up of a barista smiling at the camera, in a bright, modern coffee shop” gives the AI a clear, specific instruction on how to frame the shot, which is more effective than just “a barista in a coffee shop.”

Prompting for Realistic Portraits

A common pitfall in AI portraiture is the “uncanny valley,” where faces appear unnaturally smooth, symmetrical, and devoid of life. This is often described as the “plastic face” or “doll-like” problem.

This issue often arises from the training data of most models, which can be saturated with heavily retouched, stylized, and idealized photographs. While some models, such as Lucid Origin, are fine-tuned towards realistic results, many AIs learn to replicate these idealized patterns, over-smoothing skin and erasing the subtle imperfections that define a real human face.

Regardless of your starting model, to get the most natural output, it’s best to counteract this tendency by explicitly guiding the model toward a more authentic and imperfect representation of reality:

1. Prompt for natural skin details. Instead of letting the model default to an airbrushed look, explicitly include keywords that describe real skin texture. This forces the model to pull from more natural-looking data.
   Example terms: detailed skin texture, visible pores, freckles, slight blemishes, laugh lines around the eyes.
2. Add a “catchlight” to the eyes. A key technique in photography and cinematography to make a subject appear more alive is to ensure there is a “catchlight”—a small reflection of a light source in their eyes. Prompts without this can result in “dead” or flat-looking eyes.
   Example terms: catchlight in eyes, sparkle in eyes, light reflecting in eyes.

Let’s see this in practice and start with a prompt that is less likely to give us a realistic portrait:

A photorealistic close-up portrait of a woman with red hair, in soft lighting.

Let’s now improve the prompt by adding a catchlight and natural skin details:

A photorealistic close-up portrait of a woman with red hair, with a light dusting of freckles across her nose and cheeks, and detailed, visible skin pores, cinematic catchlight.

Prompting for Images with Text

Rendering coherent, legible, and correctly spelled text remains one of the most significant challenges for most text-to-image AI models. This difficulty arises because models are trained to learn from the pixel patterns of images, not from the symbolic and linguistic rules of language.

To an image generation model, text is not a sequence of characters forming a word but rather an intricate texture. This is why generated text often appears as gibberish or distorted, letter-like forms.

While challenging, newer models have demonstrated improved capabilities for text generation. Success, however, depends heavily on a structured prompting approach. Inside the Leonardo.Ai app, we recommend using the Ideogram 3.0 model alongside these best practices:

• Use quotation marks:
  Enclose the exact text you want to render in double quotation marks. This syntax signals to the model that the enclosed string should be treated as literal text.
• Keep text short: For optimal results, limit your text to a few words (ideally under 25 characters). The model is far more likely to render a single word or a short phrase correctly than a full sentence.
• Guide the font style: You can influence the aesthetic by specifying a general font style, but do not expect the model to replicate a specific, named font like Helvetica. Instead, use descriptive terms, like “in a clean, bold, sans-serif font” or “in an elegant script font.”
• Iterate on placement: While you can suggest placement (e.g., “in the center,” “as a title”), the AI’s control over text position is sometimes imprecise. Achieving perfect alignment may require multiple regenerations or post-processing.
• Avoid complex typography: Long sentences, paragraphs, or complex typographic layouts are highly likely to fail.

Let’s say we want to create a design for a T-shirt for a video editor, and we want to see how the text “Caffeine & Keyframes” would look:

A photorealistic mockup of a heather-gray t-shirt. On the chest, the text “Caffeine & Keyframes” is printed in a retro, slightly distressed, all-caps font. The text is flanked by a small, stylized coffee bean icon on the left and a keyframe diamond icon on the right.

Negative Prompting in Images

When an image isn’t right, the first instinct is often to try and remove the problem element. However, AI models are built to understand what things are far better than what they are not.

Telling a model to create “a room without furniture” can be confusing. It forces the AI to first generate the concept of “furniture” and then try to negate it, which can lead to artifacts or failure. Instead, we could tell the model to create “an empty room.”

Instead of negating a concept, simply describe the state you do want to see.

When You Should Use a Negative Prompt

Negative prompting can be useful when trying to remove artifacts, minor imperfections, or enforce a specific, difficult style. Let’s see an example of how we can use negative prompting to improve an image. We’ll start with this prompt:

A professional product shot of a minimalist black baseball cap. The word “Origin” is embroidered in the center in a clean, white, elegant script font. Studio lighting on a plain gray background.

On the upper side of the cap, there are some noticeable folds. It’s natural for a product to have some folds—it adds realism—but we want to reduce them for a cleaner look. We can do this using negative prompting.

With the Ideogram 3.0 model selected, click the Advanced Settings panel and toggle on Negative Prompt. This will enable you to add the negative prompt in the field below the main prompt.

Below, we see the result after adding the negative prompt “wrinkles, creases, folds, fabric imperfections, stitching errors, uneven texture” (note: we used a fixed seed to get the same cap):

If you look very, very closely at the left side of the cap, there’s still a minor fold. This illustrates an important point: negative prompts are soft constraints, not absolute commands. They assign lower weights to the unwanted terms, making them less likely to appear. This is why we always recommend seeing if it’s possible to use a positive prompt instead.

We encourage you to try generating the same image using this positive prompt on the Leonardo.Ai app (we recommend using the Ideogram 3.0 model for accurate text rendering):

A professional product shot of a minimalist black baseball cap with smooth, taut fabric and a perfectly shaped, structured crown. The word ‘ORIGIN’ is embroidered in the center in a clean, white, elegant script font. Studio lighting on a plain gray background.

When you must use negative prompting, follow these best practices:

• Be specific. Vague exclusions like “bad quality” or “weird” are ineffective. Instead, use precise terms that describe the flaw you want to avoid, such as “extra fingers,” “blurry face,” “low resolution,” or “watermark.”
• Use synonyms for reinforcement. To strengthen the model’s avoidance of a flaw, use a string of related terms. For example, instead of just “bad hands,” a more effective prompt might be “deformed hands,” “unnatural fingers,” “disfigured palms,” or “incorrect hand anatomy.”
• Don’t use instructive language. Avoid words like “no” or “don’t.” If you want to avoid greenery, your negative prompt should be “greenery,” “plants,” “forest,“ “trees,“ not “no greenery.”
• Don’t use conflicting terms. Ensure your negative prompt doesn’t contradict your positive prompt. If you ask for a “dark, moody scene,” don’t add “dark shadows” to the negative prompt.

Advanced Image Prompting Techniques

With the essentials covered, you’re already well-equipped to generate great images. To truly take your work to the next level, it’s helpful to focus less on the form of your prompt and more on its content. This means learning the specific vocabulary of visual language.

The techniques in this section are designed to give you that vocabulary, providing the granular control used by professional photographers and art directors. While there are many visual factors we can control, we’ll focus on the ones that are most useful in practice.

Camera Angle

Camera angle refers to the camera’s position relative to the subject, and it dramatically changes how the subject is perceived.

• Low-angle shot: Makes the subject appear powerful, dominant, or heroic.
• High-angle shot: Can make the subject seem smaller, vulnerable, or submissive.
• Dutch angle: Tilts the camera, creating a sense of unease, tension, or disorientation.
• Bird’s-eye view: A shot from directly overhead, offering a strategic or map-like perspective.

Portrait Lighting

Lighting is one of the most powerful tools for defining an image’s mood, dimension, and narrative tone. Here, we focus on the visual vocabulary you can use for portraits—here are the most popular studio lighting techniques:

• Rembrandt lighting: The key light is placed high and to one side of the subject (typically at a 45-degree angle). Creates a small, inverted triangle of light on the cheek opposite the light source. It’s known for being dramatic and moody.
• Butterfly lighting: The key light is placed high and directly in front of the subject, angled down. Creates a small, butterfly-shaped shadow directly under the nose. This is often seen as glamorous and flattering.
• Side lighting: The key light is placed directly to one side of the subject, at a 90-degree angle. This splits the face in half, with one side brightly lit and the other in deep shadow. This dramatic look is often used to create mystery, inner conflict, or a “double-faced” personality.
• Directional lighting: This can be placed directly below the subject, angled up, or high above the subject, angled down. Lighting from below casts unnatural, upward-pointing shadows to look sinister or frightening. Lighting from above (top lighting) creates deep shadows in the eye sockets, making a subject look mysterious or harsh.

Composition

Finally, beyond the camera and lighting, you can direct the arrangement of elements within the frame. Explicitly prompting for established compositional rules can improve the balance and visual flow of your output. Here are some of the most popular composition rules:

• Rule of thirds: Place key elements on the lines or intersections of a 3×3 grid, not the center, for a more dynamic composition. Prompt with phrases like “rule of thirds composition” or “subject in the left third of the frame.”
• Leading lines: Use lines in the scene (like roads or rivers) to guide the viewer’s eye to the main subject. Prompt with phrases like “a winding path leading to a castle” or “converging lines.”
• Framing: Use foreground elements like doorways, windows, or branches to create a “frame” around your subject. This adds depth and draws attention. Prompt by describing the frame, like “a castle framed by a stone archway.”
• Depth: Create a 3D feel by including distinct foreground, middleground, and background layers. Prompt by layering your description, such as “a foreground flower, a middleground person, and background mountains.”

Final Thoughts and Bonus Tip

AI is a powerful tool, but it relies on your direction to produce a high-fidelity result. These techniques are designed to give you that control, putting you firmly in the creator’s seat. The model provides the pixels, but you provide the vision, the context, and the creative intent.

As a bonus tip for your workflow, you can use this guide to empower your favorite chatbot, whether that’s ChatGPT, Gemini, or any AI assistant. Try pasting the key principles, best practices, and examples from this article directly into your favorite chatbot. Once it has the context of what makes a good prompt, you can give it a simple idea and ask it to write an advanced, detailed prompt for you.If you need some inspiration, head to the Community Creations section and see how others are prompting for the top-looking images. We can’t wait to see what you create!

Your next great image is one prompt away

Mastered the theory – now bring it to life with Leonardo’s models and tools.

Get started for free