Generate / Edit an Image via BlockEden Gemini Proxy

Example: text-only generation (Imagen 3)

python /path/to/skill/scripts/generate_image.py \
  "a cozy coffee shop on a rainy day, impressionist painting style" \
  --aspect-ratio 16:9 \
  --output-dir ~/Desktop

Example: reference image / editing (Gemini Flash)

python /path/to/skill/scripts/generate_image.py \
  "add a dragon flying in the background" \
  --reference-image ./my-photo.jpg \
  --output-dir ~/Desktop

Example: text-to-image with Gemini Flash

python /path/to/skill/scripts/generate_image.py \
  "a futuristic cityscape at dusk" \
  --model gemini-2.0-flash-preview-image-generation \
  --output-dir ~/Desktop

After generation

Text-only (Imagen 3) output

{
  "filepath": "/path/to/image_1234567890.png",
  "model": "imagen-3.0-generate-002",
  "aspect_ratio": "1:1",
  "prompt": "..."
}

Reference image / edit (Gemini Flash) output

{
  "filepath": "/path/to/image_1234567890.png",
  "model": "gemini-2.0-flash-preview-image-generation",
  "reference_image": "./my-photo.jpg",
  "prompt": "add a dragon flying in the background"
}

Once the image is saved:

1. Tell the user the file path so they can open it.
2. Offer to regenerate with different parameters if the user wants to iterate.

Interpreting the user's request

• Extract the core visual description from what the user says and use it as the prompt.
• If the user provides a local image file path to edit or use as reference, pass it via --reference-image. This automatically uses Gemini Flash for editing.
• If the user specifies a wide/landscape aspect, map to --aspect-ratio 16:9. Tall/portrait → 9:16. Square → 1:1.
• If the user mentions "Gemini" explicitly, use --model gemini-2.0-flash-preview-image-generation.
• If no output location is given, save to the current working directory.
• The reference image must be a local file path (not a URL). If the user gives a URL, download it first with curl -L -o /tmp/ref_image.jpg "<url>".

Prompt crafting tips

Good image prompts are specific and visual. If the user's request is vague, enhance it slightly:

• Add a medium/style if none given (e.g., "digital art", "watercolor", "photorealistic")
• Add lighting or atmosphere if it feels generic (e.g., "golden hour lighting", "moody atmosphere")
• Don't over-engineer — keep the user's intent central