Gemini 2.5 Flash Image guide to photo prompts and editing

Gemini 2.5 Flash Image is Google's hosted model for creating and editing images using text and pictures, also known as Nano Banana. Developers access it through the Gemini API, Google AI Studio, or Vertex AI; photo prompting is a way to use that model. The cited access pages provide no open-weight download for local installation. Google's launch announcement

This guide turns a photographic idea into a brief you can evaluate. The example prompts below are original exercises, not reported test results or guaranteed recipes.

What are the key facts about Gemini 2.5 Flash Image?

How do you write a photo prompt for Gemini?

Google's prompting guide describes scene-based instructions, conversational refinement, and composition from reference images. Its photographic guidance emphasizes viewpoint, lighting, and visible detail. Use these as a vocabulary for expressing your intent, then assess the actual output against that intent. Official prompting guide

Start by deciding what the viewer should notice first. A product photograph might emphasize the shape of a ceramic jug, while a portrait might emphasize expression. Write the primary subject before the background so your own brief remains easy to review.

Next, identify the light source. Instead of requesting every attractive lighting adjective, choose a coherent setup: a window beside the subject, an overcast sky, or a lamp illuminating a table. Specify the direction and softness you want to see.

For a broader explanation of Google's image products, read the sibling Nano Banana Pro overview. Keep product selection separate from prompt revision so that changing a model does not obscure whether your wording improved.

What are the limits of Gemini photo editing?

Google's launch article identifies long passages of text, character consistency, and fine factual detail as areas needing improvement. A request to preserve a face or label is therefore an instruction to evaluate, not proof that the output will preserve it. Documented development areas

The dedicated model page lists image and text as inputs and outputs, with no support for audio generation or Search grounding. A photo-style prompt does not add those capabilities. Capability table

Avoid describing the generated result as evidence of a real photographic event. In your working notes, distinguish a synthetic scene from a photograph that supplied reference material. When editing a reference, retain the original so reviewers can inspect changes without relying on memory.

Camera terminology should express visual intent. Treat a requested lens or exposure setting as part of the brief, and judge perspective, blur, and brightness in the resulting picture. Do not infer a real camera setup from the words you typed.

How do you edit a reference photo with Gemini?

Choose Gemini 2.5 Flash Image in Google AI Studio for a model-specific experiment, or call its stable API identifier, gemini-2.5-flash-image. The model documentation distinguishes that stable name from its deprecated preview identifier. Model documentation

Build your first prompt from subject, setting, framing, light, and the details you want preserved. Try this original exercise:

Create a product photograph of a handmade blue ceramic jug on a pale wooden shelf. Show the full jug from slightly above shelf height. Use soft window light from the left and a plain warm gray wall. Keep the glaze irregular and the handle fully visible.

Before generating, write down what would count as failure: a cropped handle, a different vessel shape, unwanted lettering, or distracting props. That small list gives you a repeatable way to compare variations.

For editing, use an actual reference file and describe the requested change. Install the official Python SDK and Pillow with python -m pip install google-genai pillow, create an API key, and set GEMINI_API_KEY. Configure paid-tier billing for the Developer API; this model has no free API tier. The example adapts Google's image-input pattern and saves returned image parts. Getting started Pricing Pillow installation Image-input example

from google import genai
from PIL import Image
from io import BytesIO

client = genai.Client()
response = client.models.generate_content(
    model="gemini-2.5-flash-image",
    contents=[
        "Replace the wall with pale green plaster. Keep the jug and shelf unchanged.",
        Image.open("reference.png"),
    ],
)
saved = 0
for candidate in response.candidates or []:
    if candidate.content is None:
        continue
    for part in candidate.content.parts or []:
        data = part.inline_data
        if data and data.data and (data.mime_type or "").startswith("image/"):
            Image.open(BytesIO(data.data)).save(f"edit-{saved}.png")
            saved += 1
if saved == 0:
    raise RuntimeError("No image returned; inspect the response feedback.")

The example stops if it saves no image. Inspect prompt_feedback and candidate finish reasons before retrying; the API documents these response fields. Keep the source file and save revisions separately. Response reference

For a second exercise, change the photographic category. Describe a quiet portrait of a fictional adult baker beside a workbench, with flour on an apron and indirect morning light. Specify the expression and crop, then review whether those choices support the story you intended.

For a third exercise, edit only the environment of an existing image. Ask for the background to become an uncluttered studio wall while preserving the subject's pose and clothing. Compare the subject directly with the input before deciding whether the background change succeeded.

Finally, revise a single instruction at a time. If the image feels too staged, identify the visible feature causing that impression: rigid posture, evenly arranged props, or overly smooth surfaces. Replace a vague request for realism with a concrete instruction addressing that feature.

Use the PromptZone prompt library to collect more starting ideas. Keep your accepted examples with their references and review notes so that the collection records what worked for your task.

How does Gemini photo editing compare with SDXL?

Gemini 2.5 Flash Image offers a hosted reference-editing workflow. Stability AI's SDXL Base model card supplies downloadable weights and an inference example for a different operating model. Google model page SDXL model card

For readers considering checkpoints and local workflow choices, the SDXL models guide provides the relevant pillar. Compare your ability to preserve the required subject and complete the edit, rather than assuming a prompt transfers unchanged between systems.

What else should you know about Gemini 2.5 Flash Image?

Which Gemini model should I select for these photo prompts?

Select gemini-2.5-flash-image to reproduce the model choice in this guide. Google lists that stable identifier separately from the deprecated preview version. Model documentation

Should I write a sentence or a list of keywords?

Google recommends describing a coherent scene instead of relying on disconnected keywords. Start with a short photographic brief and add details that make the intended result easier to evaluate. Prompting guidance

How do I keep the same person or product?

Provide a reference and specify the features that must remain unchanged. Inspect the output carefully because Google identifies consistency and fine detail as continuing areas of improvement. Launch announcement

Can I edit a picture with an ordinary-language instruction?

Yes, image-plus-text editing is a documented workflow. Name both the change and the parts you want preserved, then compare the saved result with the input. Prompting guidance

Sources

• Google's Gemini 2.5 Flash Image introduction
• Gemini 2.5 Flash Image model documentation
• Google's image prompting guide
• Gemini API getting started
• Gemini content-generation response reference
• Pillow installation instructions
• Gemini Developer API pricing
• Stability AI SDXL Base model card

Related guides on PromptZone

• The Ultimate Guide to Fooocus Image Prompts
• Varying Prompt Weight with Stable Diffusion