An AI image tool may give you two ways to begin: describe a scene in words, or upload an image and ask for a change. Those options look similar, but they solve different problems. If you choose the wrong starting point, you may spend several rounds trying to recreate something you already had or trying to preserve details that were never clearly defined. Nano Banana supports both text-led and image-led creation, making the choice especially relevant. The simplest rule is this: use text when you need invention, and use a reference image when you need continuity.
Use a Text Prompt When the Scene Is Still Open
Text generation works best when there is no existing image you need to protect.
Suppose you need a header image for an article about remote work. You know the message, but you do not care about a specific person, room, desk, or camera angle. A text prompt gives the model room to construct the scene from scratch.
Start with the visible essentials: who or what is in the image, where the scene takes place, what is happening, how the camera sees it, and what kind of light is present. Then add only the constraints that matter to the final placement.
For example: “A freelance designer working at a dining table in a small apartment, laptop open, late-afternoon daylight, realistic photography, medium-wide composition, clean space on the right for text.”
That is usually more useful than a prompt filled with style labels but no concrete scene.
Use a Reference Image When Something Already Needs to Stay
A reference image changes the task from invention to controlled transformation.
Imagine you already have a product photo with the correct bottle shape and camera angle, but the background looks plain. Rebuilding the whole image from text creates unnecessary risk. Instead, use the photo as the starting point and ask for a new setting while clearly stating what must remain unchanged.
The same logic applies to a portrait. If the person, hairstyle, clothing, or pose matters, the source image carries information that would be difficult to reproduce accurately from a description alone.
With Nano Banana AI, image-based generation can be used when an existing visual should guide the result. The prompt then becomes an edit brief: preserve these elements, change these elements, and keep the final composition suitable for its intended use.
A Quick Decision Table
| Situation | Better starting point | Why |
| You only have an idea | Text prompt | Nothing specific needs to be preserved |
| You need a new setting for an existing photo | Reference image | The subject and composition already exist |
| You want several concepts for a campaign | Text prompt | Wider exploration is useful |
| You need the same person or product across variations | Reference image | Continuity matters |
| You want to replace one element | Reference image | The change is narrow and defined |
| You are unsure what the scene should look like | Text prompt | The model can help explore directions |
The table is not a hard rule. Sometimes a text-generated concept becomes the reference for the next round. In practice, strong workflows often move from open exploration to more controlled editing.
Three Questions to Ask Before You Generate
When the choice is not obvious, these questions usually make it clear.
- What Must Stay Exactly the Same?
List the non-negotiable elements. It may be a face, product shape, logo placement, room layout, pose, or camera perspective.
If the list is long, use a reference image. If almost nothing must stay, text generation gives you more freedom.
This question also improves the prompt. Instead of writing “make a similar photo,” you can say, “Keep the person, jacket, pose, and camera angle unchanged. Replace only the outdoor background with a modern train station.”
- How Much Do I Want the Model to Invent?
Sometimes invention is the point. A creator may want six visual concepts for “a quiet futuristic library” without knowing what the best version looks like.
Other times, invention is a problem. If you are editing a real product photo, you do not want the model redesigning the packaging or adding accessories.
Decide where creativity is welcome and where it is not. Text prompts usually allow a wider search. Reference-led edits are better when the creative space should be narrower.
- What Will the Final Image Be Used For?
A blog illustration can tolerate more interpretation than an ecommerce image or an identity-sensitive portrait.
If the final placement depends on factual accuracy, preserve more real source material and inspect the output more carefully. If the image is conceptual, you can allow more freedom.
Usage also affects composition. A hero banner may need negative space for copy, while a square social post needs a strong central subject. Include that requirement before generation rather than fixing it later.
Write the Prompt for the Starting Point You Chose
For Text-Only Generation
A practical text prompt can follow this order:
Subject → setting → action → composition → lighting → important constraints.
That structure keeps the scene readable. You do not need to describe every surface, color, and emotion.
Avoid vague combinations such as “professional, creative, modern, premium, cinematic.” Replace them with visible choices. “Soft window light, neutral office, eye-level camera, uncluttered desk” tells the model what those qualities should look like.
If the first result misses the goal, identify the largest error. Change the framing if the subject is too small. Change the setting if the context is wrong. Do not rewrite every line at once. Controlled revisions make it easier to learn what is working.
For Reference-Image Editing
Reference-image prompts should make the boundary between “keep” and “change” obvious.
A useful pattern is: “Keep A, B, and C unchanged. Change D. Match E.”
For example: “Keep the woman’s facial features, hairstyle, black coat, and standing pose unchanged. Replace the background with a rainy city street at night. Match the original camera angle and realistic lighting.”
If you need several changes, group them logically. First state identity or product details that must stay. Then describe the new environment. Finally mention composition or lighting.
Do not assume the model knows which part of the original matters most to you. If the exact bag, chair, label, or hairstyle is important, name it. Clear preservation instructions are as important as the requested edit.
When to Combine Both Approaches
The strongest process is often not text versus image. It is text first, image second.
You might start with text to explore three directions for a campaign. Once one composition works, use that result as a reference and create controlled variations for different placements. That turns an open-ended search into a more stable series.
The reverse can also work. Begin with a real photo, edit the setting, then use the approved result as a reference for related visuals.
Think of text generation as a way to open the search space and reference images as a way to narrow it. Moving between them deliberately gives you more control than treating every generation as a fresh start.
Conclusion
Choosing the right input can save more effort than writing a longer prompt. Start from text when you want the model to invent the scene. Start from a reference image when important details already exist and should remain recognizable. Then make your constraints explicit, review the output for accuracy, and change one major variable at a time. For your next image task, ask one question before typing anything: am I creating something new, or changing something I already have? The answer usually tells you where to begin.
