Writing image prompts for GPT Image, Nano Banana and Seedream
One prompt structure that works across models, then what actually differs between GPT Image 2.5, Nano Banana 2 and Seedream in FlexPic: controls, references, resolution and price.
The three image model families you will use most in FlexPic are GPT Image (OpenAI), Nano Banana (Google) and Seedream (ByteDance). They all understand plain English, so one good prompt structure carries across them. What differs is the controls around the prompt, how references are handled and what a picture costs. This article covers both.
A prompt structure that works everywhere
Write a prompt as a short description of a photograph or illustration, in this order:
- Subject. What is in the picture, with two or three specific details.
- Setting. Where it is and what surrounds it.
- Composition. Close-up, wide shot, from above, centred, rule of thirds.
- Light. Time of day and quality: soft window light, hard noon sun, neon at night.
- Style. Photograph, flat vector illustration, watercolour, 3D render, film still.
An example you can paste:
A hand-thrown ceramic teapot on a linen tablecloth, a sprig of rosemary beside it, shot from slightly above, soft window light from the left, muted earth tones, editorial food photographyThree habits keep results predictable:
- Say what you want, not what you do not want. "A clean white studio background" works better than "no clutter".
- Be specific about count and position. "Two cups on the left, one plate on the right" beats "some cups and a plate".
- Put text in quotation marks if you want words in the picture, and keep it short:
a sign that reads "OPEN". Long paragraphs of text in an image are unreliable on every model.
Where the models differ in FlexPic
The table summarises what you control and what you pay. Prices are for one image, computed from FlexPic's price table when this page was built.
| GPT Image 2.5 | Nano Banana 2 | Seedream 4.0 | |
|---|---|---|---|
| Maker | OpenAI | ByteDance | |
| What changes the price | Quality (Low, Medium, High) and resolution | Resolution (1K, 2K, 4K) | Neither: one price at every size |
| 1K price | 8 (Low), 17 (Medium), 67 (High) | 84 | 38 |
| Higher resolution | 2K is about 23 and 4K about 45 at Low quality | 2K is 127, 4K is 189 | 38 at every size |
| Reference images | Yes | Yes | See the Model panel |
| Extra setting | Mode | None | None |
Two practical conclusions follow from the table:
- GPT Image 2.5 is the cheapest place to iterate. At Low quality a 1K image costs about 8 credits, so you can try many prompts. Raise the quality for a favourite.
- Nano Banana 2 charges by resolution, not quality. If you want a 4K result from it, expect the price to rise accordingly.
Model availability changes. The Model row in Create Image lists the models you can use right now, and a model marked Soon cannot be selected. Nano Banana 2 replaces the older Nano Banana, which Google lists as deprecated.
Using reference images
Create Image has an Add visual references area under the prompt. Upload a picture to guide the subject, style or composition, and mention what to take from it in the prompt: "Use the colour palette of the reference" or "keep the same bottle shape". Reference images are the foundation of consistent characters across a set.
Keep references clean: one subject, good light and no heavy text overlays. If you give the model contradictory information, such as a red dress in the reference and a blue dress in the prompt, expect it to pick one unpredictably. Be explicit about which wins.
Editing instead of regenerating
If a picture is nearly right, do not rewrite the whole prompt. Open Edit Image, upload it and describe only the change:
Change the tablecloth to dark blue, keep everything else exactly as it isModels can alter more than you asked. Add "keep everything else the same" and check the faces, hands and any text afterwards. Nano Banana 2 edits by description alone, with no mask, so be precise about what changes.
Prompt differences worth testing
You will get more from running one prompt on two models than from reading advice about them. A sensible test:
- Write one subject prompt with the five parts above.
- Run it on GPT Image 2.5 at Low quality: about 8 credits.
- Run it on Nano Banana 2 at 1K: about 84 credits.
- Run it on Seedream 4.0: about 38 credits.
Total for the three: 130 credits. Compare composition, how literally each followed your count and positions, text rendering and colour. Keep notes: the model that wins for product shots may not win for portraits.
Style words that travel well
- Photographic: "35mm photograph", "macro", "shallow depth of field", "overcast soft light".
- Illustration: "flat vector illustration", "risograph print", "pencil sketch", "children's book watercolour".
- 3D: "clay render", "isometric 3D scene", "glossy plastic toy".
- Graphic design: "poster layout with large headline space at the top", "square social post with a clean background".
Avoid naming living artists or copyrighted characters as a shortcut. It raises rights questions, and the responsible use guide explains why. Describe the look in words instead.
Aspect ratio and what you will use it for
Set the ratio under Output before you generate: 1:1 for profile pictures, 4:5 for feed posts, 16:9 for banners and thumbnails, 9:16 for stories. Changing it later changes the composition. The aspect ratio guide lists common targets.
When a result is wrong
If you keep getting the same mistake, the troubleshooting guide walks through the usual causes: too many subjects, conflicting instructions, vague wording and unsupported text. When the picture is ready to move, send it to Frame to Video and follow the image-to-video workflow.