Nano Banana 2

Help the user write prompts that get the image they pictured out of Google's Nano Banana models (Gemini 3 family). The models apply deep reasoning to a prompt before they render, so a narrative, well-directed prompt beats a keyword list almost every time.

• IS: prompt craft for Gemini 3 image generation and editing. Picking the right framework, structuring the prompt, and dialing in lighting, camera, color, text, and references.
• IS NOT: video prompts of any kind (Seedance video prompts are seedance), and not the structure or schema of a film's stills.json / shots.json (also seedance).

This repo can generate stills with Nano Banana: set the stills file's top-level model to a gemini-* id and vs stills routes to the Gemini image API (needs GEMINI_API_KEY). Write the still prompt with the frameworks here, place it in stills.json per seedance, then run it with vs. The frameworks also stand alone for any one-off Gemini image (posters, mockups, marketing).

Read references/frameworks.md before writing any prompt: pick the framework that matches the task. Read references/creative-direction.md to elevate a working prompt to studio quality. Read references/specs.md for model ids, aspect ratios, resolutions, limits, and API config.

The models

Three models share the Nano Banana name. Default to Nano Banana 2.

• Nano Banana 2 = gemini-3.1-flash-image. The go-to: best balance of intelligence, cost, and latency. High-volume friendly, which makes it the storyboard-batch model. Adds 0.5K resolution, extreme aspect ratios (1:4, 4:1, 1:8, 8:1), Google Image Search grounding, video-to-image, and a tunable thinking level. Up to 10 high-fidelity objects and 4 characters; no style references.
• Nano Banana Pro = gemini-3-pro-image. Professional asset production and the hardest, most complex instructions. Up to 6 high-fidelity objects, 5 characters, and 3 style references (14 reference images total), deeper default thinking, up to 4K. The best available model for legible in-frame text, which is why any film keyframe carrying signage comes from Pro.
• Nano Banana (legacy) = gemini-2.5-flash-image. Fast and cheap, 1024px, best with up to 3 input images. Use only when speed and volume beat quality.