[Image Generation] Explore Nano Banana Pro capabilities for improved visual output

[madjin]

Summary

Research into Nano Banana Pro (Gemini 3 Pro) capabilities reveals features that could benefit our poster generation pipeline. This issue tracks exploration of these capabilities.

Resources

• awesome-nanobanana-pro - Community-curated prompts and techniques
• Gemini Image Generation Docs - Official capabilities

Capabilities Worth Exploring

Feature	Potential Use Case
Google Search grounding	Accurate brand/project logos, current events imagery
Thinking mode	Complex multi-element compositions
Multi-turn editing	Iterative style refinement
Text rendering	Infographic-style posters, data visualizations
Transparent backgrounds	Cleaner icon/asset generation

Prompting Patterns to Test

From awesome-nanobanana-pro, these techniques show strong results:

1. Narrative paragraph prompts vs keyword lists
2. Camera/lens specifications for consistent rendering
3. Lighting descriptions for photorealistic outputs
4. Era-specific aesthetic triggers
5. Explicit composition guidance

Exploration Tasks

• Test Google Search grounding for brand logo accuracy
• Compare narrative vs structured prompts for our use cases
• Experiment with thinking mode for complex scenes
• Test multi-turn editing for style iteration
• Document what works best for our content (ElizaOS dev updates)

Related Files

• scripts/posters/generate-icons.py - Icon generation
• scripts/posters/generate-ai-image.py - Daily poster generation
• scripts/posters/config/style-presets.json - Existing style templates

Priority: Low - exploratory research for future improvements

[madjin]

Implementation Update: Search Grounding + Text Output

We've implemented the first set of Nano Banana Pro capabilities:

✅ Implemented

Google Search Grounding

• Enabled via plugins: [{"id": "web"}] in API requests
• Model can now look up actual brand logos before generating
• Use --no-search flag to disable if needed

Text + Image Output

• generate_image() now returns (image_bytes, text_response) tuple
• Model provides descriptions/confidence about generated icons
• Text responses saved to _grid_{type}_latest.txt for post-processing insights

🔜 Future Exploration

Thinking Mode (include_thoughts=True)

• Could improve planning for complex grid layouts
• Access model's reasoning for better debugging
• Useful for understanding why certain logos were generated a certain way

Multi-turn Chat Mode (client.chats.create())

• Enable iterative refinement: "fix cell 3, the Discord logo is wrong"
• Maintain character/style consistency across edits
• Could be powerful for fixing specific icons without regenerating the whole grid

Higher Resolutions (2K/4K)

• image_size: "2K" or image_size: "4K" in config
• Useful for high-quality final exports
• 4K is more expensive, use sparingly

Usage

# With search grounding (default)
python scripts/posters/generate-icons.py --missing --type project

# Without search (faster, cheaper)
python scripts/posters/generate-icons.py --missing --no-search

Reference: Nano Banana Pro Notebook