Since finishing our first text-to-image prototype, we’ve generated thousands of images and learned a lot about how text-to-image works. Many of those lessons have already been applied to the UX and the style presets of BeCasso’s newest addition, helping you write prompts easily and create beautiful images fast. However, some things – like writing a good prompt or controlling specific aspects of a scene such as the camera position – are hard to put into a UI. That’s why we want to share our best practices and lessons learned with you today.
Anatomy of a Good Prompt
BeCasso’s text-to-image feature is straightforward: write a prompt and click generate to create your first image. This works perfectly well when your prompt already contains everything the AI needs to bring your idea to life. But what makes a good prompt?
According to https://stable-diffusion-art.com/prompt-guide/ a good prompt consists of the following building blocks:
1. Subject:
The subject basically describes the main thing you want to see in the image. This can be as simple as “A bonsai tree” or a more detailed description of the scene. Since text-to-image has a limit on prompt length, the art is in being descriptive yet concise — every word should have an impact on the result.
2. Detailing the subject:
The less you specify, the more the AI will fill in the gaps on its own. Writing short, descriptive prompts that define the most important aspects of your subject is the key to great results. Let’s add more detail to our subject: “A maple bonsai standing in a Japanese garden.” Although no color is specified, the AI adds red leaves, since most images of maple bonsai on the internet have reddish leaves.
4. Detailing the Style:
While the medium defines the broad category, style provides additional detail. For example, you can choose oil painting as the medium and add “impressionist” or “surrealist” as a style. Following the Oil Masterpieces preset, we add “chiaroscuro” and “Dutch Golden Age.”
6. Color:
You can add colors to your prompt, but controlling them precisely is mostly a gamble. The AI tends to interpret colors as general guidance for the entire scene rather than applying them to specific objects. In our example, adding “yellow leaves” did not yield any difference, most likely because the style-related keywords had already defined the overall color mood. Emphasizing the leaf color with “bright yellow leaves” actually made an impact.
3. Defining the Medium:
Medium describes the “material” your image should resemble, such as photography, oil painting, or 3D rendering. We add the medium-related keywords to the style prompt field to keep the subject prompt clean. For our example, we borrow “classical oil painting, museum quality, rich textures” from the Oil Masterpieces preset.
5. Adding Inspiration:
This might sound surprising, but adding the name of an art-sharing website such as “DeviantArt” for 3D renderings, or an artist name such as “Annie Leibovitz” for photos, can make a noticeable difference to the output. For our example, we go with an artist from the Dutch Golden Age and add “Jan van Goyen” to the prompt. During experimentation, we learned that picking an artist who focused on landscape paintings worked much better than choosing one known for portraits.
7. Lighting:
This is especially relevant when using photography as your medium. Terms like “studio lighting” or “backlighting” can have a huge impact on the final look. This works best for images that resemble a photo, but even for our oil painting example, adding “high key lighting” definitely impacted the result.
5. Adding Inspiration:
This might sound surprising, but adding the name of an art-sharing website such as “DeviantArt” for 3D renderings, or an artist name such as “Annie Leibovitz” for photos, can make a noticeable difference to the output. For our example, we go with an artist from the Dutch Golden Age and add “Jan van Goyen” to the prompt. During experimentation, we learned that picking an artist who focused on landscape paintings worked much better than choosing one known for portraits.
6. Color:
You can add colors to your prompt, but controlling them precisely is mostly a gamble. The AI tends to interpret colors as general guidance for the entire scene rather than applying them to specific objects. In our example, adding “yellow leaves” did not yield any difference, most likely because the style-related keywords had already defined the overall color mood. Emphasizing the leaf color with “bright yellow leaves” actually made an impact.
7. Lighting:
This is especially relevant when using photography as your medium. Terms like “studio lighting” or “backlighting” can have a huge impact on the final look. This works best for images that resemble a photo, but even for our oil painting example, adding “high key lighting” definitely impacted the result.
2. Detailing the subject:
The less you specify, the more the AI will fill in the gaps on its own. Writing short, descriptive prompts that define the most important aspects of your subject is the key to great results. Let’s add more detail to our subject: “A maple bonsai standing in a Japanese garden.” Although no color is specified, the AI adds red leaves, since most images of maple bonsai on the internet have reddish leaves.
3. Defining the Medium:
Medium describes the “material” your image should resemble, such as photography, oil painting, or 3D rendering. We add the medium-related keywords to the style prompt field to keep the subject prompt clean. For our example, we borrow “classical oil painting, museum quality, rich textures” from the Oil Masterpieces preset.
4. Detailing the Style:
While the medium defines the broad category, style provides additional detail. For example, you can choose oil painting as the medium and add “impressionist” or “surrealist” as a style. Following the Oil Masterpieces preset, we add “chiaroscuro” and “Dutch Golden Age.”
Making short prompts look Awesome
As we’ve seen, writing the perfect prompt is nearly impossible and takes practice. That’s why we decided to split the “What” (your subject) from the “How” (everything else) and help you with the how.
BeCasso comes with more than 80 style presets that help you create beautiful images, even with short prompts. You can focus on what you want to see and easily explore different looks by browsing through the presets. We’ve made the style prompt visible so you can learn from it and adapt it for your own creations.
You might notice that styles like “Van Gogh Wheatfield” produce particularly authentic results despite having a rather short style prompt. The reason is that we’ve already sneaked in a future feature: the ability to provide an additional style image.
You are as old as you feel
One issue we noticed with text-to-image is that it can be surprisingly hard to control the age of a person. We tried to create an image of a “35 year old man” and were surprised when the AI generated a young child. Further experimentation showed that even “55 year” or “75 year” produced identical results. The reason is that the AI splits multi-digit numbers into individual digits and only reacts to one of them. For example, “57 year old” produces a slightly older-looking boy, because the AI picks up on the “7” rather than “57” as a whole.
The fix is simple once you know the trick: instead of a numeric age, use life-stage keywords such as “toddler”, “young child”, “boy”, “teenager”, “young adult”, “middle aged”, or “elderly”.
While this already works well, there are some limitations. One important detail: avoid hyphens in your prompts. For example, “middle-aged” should be written as “middle aged” (without the hyphen), because the AI interprets the hyphen character as a separator with special meaning, which can interfere with how your prompt is processed.
At 44, I would most likely count as a “middle aged” man – but I could hardly identify with what the AI produced. We needed to tweak it further by adding age-related visual clues like “salt and pepper hair” or “slight wrinkles”. While the result looked somewhat younger, it still wasn’t quite right.
The real trick turned out to be blending descriptors from different age categories. In the end, “mature young man” hit much closer to the mark than “middle aged man” ever did.
Camera Control
Controlling camera position and angle can be tricky, and it doesn’t work equally well for all subjects. The AI can only reproduce camera angles it has “seen” frequently in its training data. A bird’s-eye view of a bonsai tree, for instance, is an uncommon composition in photography, so the AI has little to draw from.
Camera Angle
That said, the following tips can give you some control over the camera position. While the AI has a general understanding of photography terms like “low angle” or “high angle”, it works much better when you explicitly describe both the camera position and its viewing direction. For example, “low angle view from ground level, looking up” is far more effective than just “low angle shot”. The following prompt extensions worked well in our experiments:
Camera Distance
The same principle applies to camera distance: the AI needs to have seen enough examples of a subject at different distances. Camera distance control therefore works best for common solitary subjects like people, cars, or cats.
In our example, the original prompt produced a close-up photo of a cat playing with a ball of yarn. Adding “medium shot” moved the camera slightly further away. To see the entire cat, adding “full body” to the prompt worked well. Interestingly, “wide shot” produced a similar framing but added much more depth of field, showing more of the environment.
Position Matters
While you can write prompts of any length, longer is not always better. The on-device AI has limitations on what it can process, and the maximum prompt length is indicated by the small number next to the prompt field. It shows how many tokens are left before your prompt gets truncated.
Beyond length, the position of keywords in your prompt also matters. Words placed earlier carry more weight. The following example doubles as a demonstration of a limitation of the lightweight AI model — it struggles with spatial relations — but it clearly shows how word order affects the result. Consider the prompt: “a large red cube in the center, a small blue cube next to it.” Simply swapping the two halves shifts the result from a red-dominant to a blue-dominant image.
Describe What You Want, Not What You Don’t
One of the most common mistakes when writing prompts is telling the AI what you don’t want. Phrases like “no wrinkles,” “no beard,” or “without glasses” seem intuitive, but they often backfire. The reason lies in how the AI processes text: the model’s text encoder (called CLIP) breaks your prompt into tokens and converts them into numerical representations. It does not truly understand negation. When you write “no people,” the AI still activates the concept of “people” — and may actually emphasize it in the result.
The solution is to rephrase everything in positive terms — describe the outcome you want, not the thing you want to avoid. Here are some common examples:
- Instead of “street with no people” → use “lifeless, empty street”
- Instead of “no wrinkles” → use “smooth skin” or “clear complexion”
- Instead of “no background” → use “plain white background” or “solid color background”
- Instead of “not blurry” → use “sharp focus” or “high detail”
This principle extends to more complex scenarios too. If the AI keeps generating an unwanted element — say, a hat on a character — don’t add “no hat.” Instead, try describing the hairstyle you want, which naturally crowds out the hat. Think of it this way: every word in your prompt is a positive instruction. The AI will try to include whatever concepts you mention, regardless of whether you prefix them with “no” or “without.” Use your limited token budget to describe only the things you actually want to see.
The prompt “A red modern Porsche on a street in a city with no people” still created an image with people walking along a street. Reformulating it to “A red modern Porsche on a lifeless, empty street in a city” created the desired result.
Conclusion
Translating the images in your mind into words has never been easy—whether you’re describing them to another person or guiding an AI to bring them to life. Yet that’s what makes prompting such a fascinating skill: it sits at the intersection of imagination, language, and creativity.
Throughout this guide, we’ve explored the core principles of effective prompting, along with practical examples of what works—and what doesn’t. While there are no perfect prompts, understanding these fundamentals will help you communicate your ideas more clearly and unlock more consistent, compelling results. Ultimately, the best way to improve is to experiment. Every prompt is an opportunity to refine your vision, discover new possibilities, and develop your own creative style. So start exploring, keep iterating, and don’t be afraid to surprise yourself.
If you’d like to put these techniques into practice, you can download our BeCasso app and start creating your own AI-powered artwork: https://apps.apple.com/us/app/becasso-photo-to-painting-app/id1125534178
We’d also love to hear from you. If there are additional topics you’d like us to cover, feel free to contact us at mail@digitalmasterpieces.com.
Finally, if you’re interested in a more literary perspective on prompting, we highly recommend this article by Ana María Caballero. It explores the striking parallels between poetry and prompt writing, highlighting why the careful choice of words matters just as much when creating images as it does when crafting verse: https://www.ie.edu/insights/articles/from-prompt-to-image-the-literary-implications-of-text-to-image-ai/
Happy prompting—and happy creating.
BeCasso
Clip2Comic
Graphite
Waterbrush
Oilbrush
Artcard