Google: Nano Banana 2 (Gemini 3.1 Flash Image Preview)

Generate a product photo of a minimalist ceramic coffee mug on a wooden table with soft morning light from the left. Clean background, shallow depth of field.

The model produces a photorealistic rendering of a white ceramic mug positioned slightly off-center on warm oak planks. Morning sunlight streams from the left, creating a gentle gradient across the mug's surface and casting a soft shadow to the right. The background blurs into a creamy bokeh, keeping focus tight on the mug's clean lines and subtle glaze texture. The lighting feels natural, with no harsh edges or artificial glow.

This example highlights Nano Banana 2's strength in product photography composition and natural lighting simulation. The 65k token context window allows detailed scene descriptions without truncation. However, at $3/Mtok output pricing, generating multiple variations for client review becomes expensive compared to dedicated product-shot models.

Create an editorial illustration for a tech article about distributed systems: abstract geometric shapes representing nodes in a network, connected by flowing data streams. Teal and coral color palette, isometric perspective.

The model renders a clean isometric grid populated with hexagonal and cubic nodes in graduated teal shades, connected by animated-looking coral lines that curve and branch organically. Each node features subtle internal detail suggesting activity—small dots or pulses—while the data streams have a hand-drawn quality that softens the technical subject. The composition balances geometric precision with approachable visual warmth, suitable for a magazine spread or blog header.

Demonstrates the model's ability to translate abstract technical concepts into accessible visual metaphors. The multimodal input support means you could reference existing diagrams or wireframes in your prompt. The Flash architecture delivers these illustrations quickly, though fine control over specific node positions or connection patterns may require iteration.

Design a fantasy book cover: a lone figure in a hooded cloak standing at the edge of a cliff overlooking a vast bioluminescent forest under twin moons. Painterly style, rich purples and greens.

The model generates a vertical composition with dramatic scale: a small silhouetted figure dominates the foreground against an expansive valley of glowing trees rendered in layered purples, teals, and acid greens. The twin moons hang low on the horizon, casting overlapping halos. Brushstroke texture is visible throughout, giving the piece a traditional fantasy-art feel rather than digital smoothness. The atmospheric perspective creates genuine depth, with distant trees fading into misty darkness.

Showcases stylistic range beyond photorealism—the painterly rendering and atmospheric mood align well with genre fiction requirements. The image modality input means you could provide reference art or mood boards. Trade-off: at preview stage, this model may lack the resolution or fine detail control needed for print-ready cover art without upscaling.

When you're tagging 500+ product images daily on a tight budget

A 4-person e-commerce team uploads 80-120 product shots every weekday and needs consistent alt-text, category tags, and defect flags before the images hit Shopify. Nano Banana 2 wins here because the $0.50/Mtok input rate makes high-volume image analysis economical—you're looking at roughly $2-3/day even at 500 images if prompts stay under 200 tokens. The 65k context window lets you batch 15-20 images in one call with a shared tagging schema, cutting API overhead. Trade-off: without public benchmarks you can't verify accuracy against GPT-4V or Claude 3.5 Sonnet on edge cases like fabric texture or color precision. If more than 5% of your catalog needs manual correction, test a sample batch against a benchmarked model first. For straightforward product photography where speed and cost matter more than perfection, this model closes the loop.

Why this model struggles with nuanced UI critique at any scale

A 12-person design agency wants to auto-generate first-pass feedback on Figma screenshots—flagging alignment issues, contrast problems, or missing states before human review. Nano Banana 2 isn't the right call. The lack of public benchmarks means you have no baseline for how well it detects subtle layout bugs or interprets design intent compared to models with published UI-understanding scores. At $3/Mtok output, verbose feedback on 50 screens/day costs $4-6, which is competitive—but you're paying for uncertainty. If the model misses 20% of actionable issues, your designers waste time on double-checks that erase the efficiency gain. The 65k window is helpful for multi-screen flows, but without proven accuracy on visual reasoning tasks, you're better off with Claude 3.5 Sonnet or GPT-4V where you can trust the critique quality. Use this model only if you're prototyping the workflow and plan to validate every output manually.

When Nano Banana 2 handles expense reports for under $10/month

A 6-person consulting firm processes 40-60 expense receipts weekly—mostly restaurant bills, ride-shares, and hotel invoices—and needs line-item totals, dates, and vendor names pulled into Airtable. Nano Banana 2 is a strong fit. The $0.50 input rate means each receipt costs roughly $0.002 if the image encodes to 4k tokens, so 200 receipts/month runs under $0.50 in input fees. Output is structured JSON (maybe 150 tokens per receipt), adding another $0.09/month at $3/Mtok. Total monthly cost: under $1 even with retries. The 65k context lets you process 10 receipts in one call if you want to amortize API latency. Risk: without benchmarks you don't know error rates on faded ink, crumpled paper, or non-English text. Run a 50-receipt pilot and measure how many need manual correction. If accuracy is above 90%, deploy it—you'll save 4 hours/week at negligible cost.