The fastest way to learn AI image generation is not to read fifty prompt formulas. It is to make one image, change one variable, and notice what happened. A short practice session can teach more than a long list of tricks because you see how wording, references, composition, and revision interact. Nano Banana can support both text-based generation and image-based editing, which makes it suitable for a simple study exercise. The goal of this project is not to produce a portfolio masterpiece. It is to finish with a repeatable method you can use on future creative work.
Pick One Small Project Before the Timer Starts
Choose a subject you can understand visually without research. Good options include a reading corner, breakfast table, bicycle outside a shop, houseplant beside a window, travel poster concept, or simple product still life.
Avoid a complex scene with ten characters and five actions. You are trying to learn cause and effect.
Write one sentence describing the finished image. For example: “A quiet breakfast table near a window on a rainy morning, photographed from seated eye level.” That sentence becomes the fixed idea for the session. You can change style, lighting, framing, or details later, but the basic scene stays stable enough for useful comparison.
Use a Simple 45-Minute Schedule
A timer prevents endless generating and forces you to make decisions.
| Time | Task | Goal |
| 0–5 min | Define the scene | One clear visual idea |
| 5–15 min | Generate first versions | Identify what is missing |
| 15–25 min | Revise one variable | Learn what the prompt changed |
| 25–35 min | Refine or use a reference image | Improve composition or consistency |
| 35–42 min | Compare results | Choose the strongest direction |
| 42–45 min | Write notes | Record what you learned |
Do not judge the exercise by whether the first image is good. Judge it by whether each round gives you information for the next one.
If you finish a stage early, do not immediately generate more images. Use the spare minute to name what you learned. For example, note that the top-down view hid the window, or that the wide frame made the table feel too empty. The timer is there to create reflection, not pressure.
Build the Prompt in Four Passes
Instead of writing one enormous prompt immediately, add information in stages. This makes it easier to see which words affect the result.
1. Describe the Subject and Action
Begin with the concrete scene: “A ceramic mug, toast, and an open paperback on a small wooden breakfast table.” If a person is present, describe the action clearly: “A student writes notes beside an open laptop.”
You are checking whether the system understands the objects and their relationship. If the arrangement is wrong, fix that before adding stylistic detail.
2. Add Camera and Composition
Now specify how the viewer sees the scene. Try “eye-level medium shot,” “top-down composition,” “close-up,” or “wide room view.” Add placement when necessary: “mug in the foreground, book slightly behind it, window on the left.”
A strong scene with poor framing still feels weak. Generate again and compare only composition, keeping the subject description as similar as possible.
3. Add Light and Mood
Once the composition works, add atmosphere. Try “soft overcast morning light,” “warm late-afternoon sun,” or “single desk lamp at night.”
Notice how mood words affect more than color. They may change contrast, background detail, materials, or even the implied story. If the rainy breakfast suddenly becomes gloomy when you wanted calm, revise the emotional language instead of adding more adjectives. “Quiet and cozy” may guide the result better than “dramatic, cinematic, atmospheric, emotional.”
4. Add Restrictions and Important Details
The final pass protects details that keep changing. You might add “no people,” “no visible text,” “keep the table uncluttered,” or “only one mug.”
Restrictions are most useful when they respond to an actual problem you saw. Do not begin with twenty negative instructions. Generate, observe the failure, and then write the smallest instruction that addresses it. This keeps the prompt readable and teaches you why each phrase exists.
Switch to Image-to-Image When Composition Is Already Close
Sometimes you get a version with the right camera angle and object placement but the wrong mood. Starting again from text may destroy the composition you wanted to keep.
In a learning exercise, Nano Banana AI can be used to take one promising result or your own reference image and request a targeted change.
Try a narrow instruction such as “keep the composition and objects; change the light to soft rainy morning daylight.” Then compare it with a fresh text generation. You will learn when references provide better control than additional words.
This is also a useful way to understand the difference between creating from scratch and revising an existing direction. The more clearly you can identify what already works, the easier it becomes to ask for a controlled edit rather than resetting the entire image.
Compare Images Side by Side Instead of From Memory
Save one image from each pass. At the end, put them next to one another and ask what actually improved.
Do not use “I like this one more” as the only criterion. Look separately at composition, object accuracy, lighting, mood, and unwanted changes. You may discover that version two had the best layout, while version four had the best light but introduced extra objects.
This is valuable. It shows that “better” is not one-dimensional. In future work, you may decide to return to an earlier version and refine it rather than automatically continuing from the newest output.
A small comparison sheet can also become a personal prompt reference library.
Write three notes before you finish. The last three minutes matter because learning disappears quickly when you only save images.
Write one sentence for each of these questions:
- Which instruction changed the image most?
- Which phrase produced an unwanted side effect?
- What would I do first if I repeated the same project tomorrow?
Your notes might say, “Top-down changed the scene more than the style words,” or, “Adding ‘cinematic’ made the table too dark.” Those observations are more useful than copying a stranger’s long prompt because they come from your own test.
Over time, these small records reveal patterns in how you describe scenes and where your prompts tend to become vague.
Repeat the Exercise With a Different Constraint
Do not repeat the exact same image next time. Keep the method but change the challenge.
If your first session used text only, begin the next with a phone photo and practice image-to-image changes. If you created a still life, try a simple human activity. If you worked in a square format, try a wide composition and notice how the scene changes.
Changing one major constraint keeps the exercise fresh while preserving the learning structure. After three or four sessions, you will have practiced subject description, framing, lighting, revision, references, and evaluation without needing a giant prompt handbook.
That is enough foundation to approach larger creative projects with much more control.
Conclusion
A useful AI image practice does not need to last all afternoon. Give yourself 45 minutes, hold one scene steady, and change the prompt in deliberate steps. Compare each version, write down what actually changed, and use image-to-image editing only when you already have something worth preserving. The point is not to discover a perfect prompt formula. It is to build the habit of observing cause and effect. Pick one simple scene, set a timer, and finish your first focused prompting exercise today.