Nano Banana 2: A Speed, Image Quality, and Editing Test

Nano Banana 2 is Google’s product name for Gemini 3.1 Flash Image. Google launched the model on February 26, 2026 and documents image generation, conversational editing, text rendering, search grounding, multiple reference images, and several output resolutions.

Google also describes the model with promotional quality and speed language. Those are provider claims. The previous version of this article converted them into an independent quality ranking and added unsupported adoption figures. This revision removes those claims and defines a test readers can reproduce.

Capabilities documented by Google

The current Gemini API documentation lists:

• text and image inputs with image and text outputs;
• conversational image editing;
• 0.5K, 1K, 2K, and 4K output options;
• image and web search grounding;
• additional aspect ratios;
• support for multiple object and character references; and
• SynthID and C2PA-related provenance work described in Google’s launch materials.

Exact limits and model identifiers can change. Confirm the live model documentation and selected product surface before building a workflow.

Test 1: text rendering

Create the same three assets in each candidate model:

1. a poster with a six-word headline;
2. a product label with mixed case, numbers, and punctuation; and
3. a two-column infographic with four short labels.

Record spelling, missing characters, layout, hierarchy, and whether a local edit preserves the rest of the image. Do not count a visually attractive image as a pass when the required words are wrong.

Test 2: reference fidelity

Use approved references for one person, one product, and one visual style. Change only the background or camera angle in each attempt.

Inspect face shape, clothing, product geometry, logo, colors, and small identifying details. Google’s documented reference-image limits show that the input workflow exists; they do not guarantee identity preservation for every asset.

Test 3: conversational editing

Start with one accepted image and request three isolated changes:

• replace one object;
• change one line of text; and
• adjust lighting without moving the subject.

Record whether unrelated regions change. A useful editor should preserve approved content while applying the requested local revision.

Test 4: latency and accepted-output cost

Measure from request submission to complete usable output. Keep network, account tier, resolution, number of images, and prompt constant. Run several attempts and report median and range rather than the fastest result.

Then calculate cost per accepted asset, including failed generations and reviewer time. “Flash” identifies Google’s model family and positioning; it is not a cross-vendor latency result.

Test 5: grounded informational images

When using search grounding for an infographic or current subject, inspect every label and factual element against the returned or independently opened sources. Image grounding can help supply context, but it does not transfer source authority to a generated graphic or eliminate hallucination risk.

Evidence boundary

Flowith did not run a controlled Nano Banana 2 comparison for this page. We therefore do not claim that it is faster, more photorealistic, or more consistent than another model. Google’s examples and statements are labeled as provider evidence; readers should compare actual outputs under matched conditions.

Sources

• Google DeepMind: Nano Banana 2
• Google launch announcement for Nano Banana 2
• Gemini 3.1 Flash Image model documentation
• Gemini API image-generation documentation

Reviewed by Flowith Lulu on September 9, 2026. Unsupported independent-comparison and adoption claims were removed.