ai-graphics

AI Graphics

Generate AI graphics — banners, hero images, illustrations, social cards, README artwork — using Fal AI's Nano Banana 2 model via the nano-tech-diagrams MCP server.

MCP Server Location

Always use mcp__jungle-shared__nano-tech-diagrams__* (runs on ubuntuvm).

Do not use jungle-local — it claims to accept local filesystem paths but runs in an environment that cannot read the workstation's filesystem, so every local-path call fails with ENOENT. The shared instance is the working path; pair it with the staging endpoint below for any image input.

Hard rules

• Always use resolution: "1K". Do not use 0.5K, 2K, or 4K unless the user explicitly overrides. 1K is the standard for this workflow.
• Pick an appropriate aspect_ratio for the use case:
  • Banners / hero images: 21:9 or 16:9
  • Social cards / OG images: 16:9
  • Square avatars / icons: 1:1
  • Portraits: 2:3 or 9:16
• Save to a sensible path using download_to (e.g., assets/banner.png, images/hero.png) — don't just leave the fal.media URL floating.
• Default output_format to png unless the user asks otherwise.

Prompt style

Nano Banana 2 rewards detailed, descriptive prompts. Include:

• Subject and pose
• Setting / background
• Lighting and mood
• Color palette
• Style descriptors (e.g., "polished digital illustration", "cinematic", "flat vector", "watercolor")
• Intended use (e.g., "suitable for a GitHub repository banner")

Keep prompts one dense paragraph rather than a bullet list.

Procedure: text-to-image

1. Confirm subject, vibe, aspect ratio with the user if unclear.
2. Ensure the target directory exists (mkdir -p via Bash).
3. Call mcp__jungle-shared__nano-tech-diagrams__text_to_image with the prompt, aspect ratio, and output_format.
4. The MCP returns a fal.media URL. Do not rely on the MCP's download_to parameter — the ubuntuvm process cannot write to the workstation filesystem. Instead, capture the URL and download it yourself:

   curl -sSL -o "/absolute/target/path.png" "<fal.media URL>"
   

5. Report the local path to the user.

Example:

mcp__jungle-shared__nano-tech-diagrams__text_to_image(
  prompt="<detailed paragraph>",
  aspect_ratio="21:9",
  output_format="png"
)
# then: curl -sSL -o /abs/path/banner.png <returned URL>

Procedure: image-to-image (and whiteboard_cleanup)

Local input files must be staged to S3 first — the MCP cannot read local paths directly, but it accepts http(s) URLs.

1. Confirm the transformation with the user (prompt, aspect ratio, target output path).
2. Stage the input file to the mcp-staging S3 bucket:

   URL=$(s3-stage /local/path/to/input.jpg)
   

   s3-stage returns a presigned GET URL (default 1-hour TTL).
3. Call mcp__jungle-shared__nano-tech-diagrams__image_to_image with image_path="$URL", plus your prompt, aspect ratio, and output_format. Use dictionary_words for any text that must be spelled correctly.
4. The MCP returns a fal.media URL. Download to the target folder:

   curl -sSL -o "/absolute/target/path.jpg" "<fal.media URL>"
   

5. Report the local path to the user.

Staged objects expire after 1 day (bucket lifecycle), so you don't need to clean them up.