Nano Banana 2 (Gemini 3.1 Flash Image) is not only a text-to-image endpoint—it is a multimodal editor that accepts follow-up turns. Google documents up to 14 input images per prompt and conversational edits that keep visual context across turns when using the API correctly.

Resolution tiers (pick before you prompt)

Tier	Typical use	Cost note
512px	Thumbnails, rapid layout probes	Lowest token burn
1K	Social, blog heroes	Default balance
2K	Print-adjacent marketing	Higher output tokens
4K	Product hero, fine type	Slowest; verify typography at size

Choose tier once per project. Jumping from 512 exploration straight to 4K delivery without re-checking composition wastes tokens.

Fourteen references without identity collapse

Assign roles in plain language:

Image 1–2: face and hair identity (ignore backgrounds).
Image 3: jacket fabric swatch only.
Image 4: lighting reference from a sunset plate.
Generate subject in a rainy alley, same identity, new wardrobe from Image 3.

More references ≠ more control. Cap distinct subjects at what the scene needs—often 3–5 images plus a text brief beats maxing attachments.

Try Nano Banana in MidassAI

Conversational edit turns (API mental model)

Turn 1: generate base scene. Turn 2: Change only the sign text to "OPEN". Keep architecture and rain unchanged. Turn 3: Warm color grade +10%. Do not move camera.

Each turn should change one dimension (text, wardrobe, grade, crop). Multi-request paragraphs cause the model to average conflicts.

Developers using google-genai should keep the same conversation session so Thought Signatures / visual context carry—dropping to stateless calls between turns resets composition.

Web grounding vs reference edits

Google Search grounding answers factual prompts (today’s weather in Seoul reflected in windows). Reference edits answer identity prompts (this exact person, new outfit). Mixing both in one turn is valid but spell which inputs are factual vs identity.

Enable grounding only when time-sensitive facts matter—otherwise latency rises without benefit.

Typography pass

Nano Banana 2 still rewards quoted strings for type: A neon sign reading "MIDASS" beats unquoted ALL CAPS mush. After a conversational grade change, re-verify spelling on a zoom pass—glow effects hide errors until export.

Failure modes

Issue	Fix
Face drifts on turn 3	Re-anchor with the original portrait reference each turn
Background swaps when editing text	Add `keep background pixels unchanged` explicitly
4K soft	Regenerate at 2K with sharper lens language, then upscale tier

Mini checklist

Tier chosen for delivery target
References cropped and role-labeled
One change per conversational turn
Grounding only when facts are time-sensitive

vs Nano Banana Pro

Use Pro when you need maximum single-shot fidelity on complex diagrams; use Flash Image (NB2) for tight iteration loops and multi-turn edits. Many production sets generate in NB2, then one Pro pass for hero stills only.

Prerequisites and setup

Before initiating multi-turn sessions, ensure your Google Cloud project has the Vertex AI API enabled with permissions for models.generateContent. You need a valid API key attached to a billing account supporting image generation endpoints. Verify you are calling the specific gemini-3.1-flash-image model string; older Flash variants lack the 14-image context window required for complex identity retention.

Local testing works best through Python scripts using the google-genai SDK version 0.5 or higher, as the SDK manages session state automatically. Keep environment variables secure and confirm your quota allows for high-resolution output tokens, as 4K generations consume significantly more capacity than standard thumbnail requests.

Extended prompt workflow

Initialize the chat session with temperature=0.7 to balance creativity against strict adherence to reference assets. Set max_output_tokens to 4096 to accommodate high-resolution image data without truncation.
Upload your primary identity reference as the first content part, labeling it system_instruction: "Maintain facial geometry from Image 1". Follow this with secondary texture references labeled strictly by material type, such as denim or leather.
Execute the first generation with a negative prompt constraint: no background changes, no lighting shifts. This locks the environment while the model renders the subject.
For the second turn, modify only the color_grade parameter in your prompt description. Use specific values like increase saturation by 15% rather than vague terms like make it pop.
Finalize the export by requesting the raw PNG output instead of compressed JPEG to preserve text clarity on signage and wardrobe details.

Common mistakes

Session Resetting: Closing the chat object between turns forces the model to re-infer context from scratch, causing identity drift. Fix: Persist the chat session object in memory throughout the entire editing loop.
Overloading Context: Attaching all 14 images in the first prompt confuses the attention mechanism. Fix: Introduce references progressively; add new assets only when that specific element requires modification.
Vague Modification Requests: Asking to improve the look triggers hallucinated changes across unrelated areas. Fix: Isolate the edit target using bounding box descriptions or explicit layer instructions like modify only the foreground subject.

Try this in MidassAI

You can replicate this multi-turn editing pipeline without writing code by using the visual interface in MidassAI Studio. The platform handles session persistence and reference labeling automatically, allowing you to focus on iterative design rather than API state management. Start your project here: https://www.midassai.com/studio/nano/

Nano Banana 2: Conversational Edits with 14 References