Gemini Image Generation

Generate high-quality AI images using Google's Gemini API.

Before First Use

Run the setup check to validate environment:

"${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts/check-setup.sh"

This validates Python, venv, dependencies, API key, and shows existing images.

Generation

SCRIPT_DIR="${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts"
"$SCRIPT_DIR/generate.sh" "your detailed prompt" \
  --aspect-ratio 16:9 \
  --resolution 4K \
  --output images/slug-v1/image.png \
  --user-request "user's original request verbatim" \
  --composition "your reasoning: why you chose this style, composition, colors, etc."

Options:

• --aspect-ratio: 1:1, 3:4, 4:3, 2:3, 3:2, 16:9, 9:16, 21:9, 9:21, 32:9, 2:1 (default: 1:1)
• --resolution: 1K, 2K, 4K (default: 2K)
• --output: Output filename (default: output.png)
• --fast: Use faster model (lower quality)
• --images: Reference image(s) for editing/fusion
• --user-request: Original user request (ALWAYS pass this)
• --composition: Your reasoning/composition notes explaining prompt choices (ALWAYS pass this)
• --no-metadata: Skip saving metadata YAML file

Output: The script saves {image_name}_metadata.yaml alongside the image containing: user request, composition reasoning, final prompt, all parameters, token usage, timestamps, and model response.

Invocation Modes

Interactive (no prompt or vague prompt): Use AskUserQuestion to gather subject, style, aspect ratio, quality.

Direct (detailed prompt provided): Execute immediately with sensible defaults.

Programmatic (called from workflow): Never block, use defaults, execute immediately.

Workflow

1. Check setup: Run check-setup.sh script (especially on first use or errors)
2. Determine mode: Interactive vs direct based on prompt clarity
3. Capture user request: Save the user's original request verbatim for --user-request
4. Compose prompt: Enhance the request into a detailed prompt, document your reasoning for --composition
5. Generate slug: 2-4 word description, lowercase, hyphenated (e.g., sunset-mountains)
6. Check existing: Look for images/{slug}-v* folders for iterations
7. Generate: Run script with ALL context flags (--user-request, --composition)
8. Report: Show user the result location and version (metadata auto-saved as image_metadata.yaml)

Interactive Questions (when prompt is vague)

When the user's request lacks detail, use AskUserQuestion to gather:

• Subject: Person/Portrait, Product, Landscape/Scene, or Abstract/Artistic
• Purpose: Final/Production (4K), Testing/Draft (fast mode), or Web/Social (2K)
• Style details: Lighting, colors, mood if relevant

Iteration Detection

User wants iteration when they say: "another version", "adjust", "modify", "change", "try with", "regenerate".

Find latest version, increment, archive previous in archive/v{N}/.

Transparent Background Requests

Gemini cannot generate transparent images. All generated images have a solid background.

When the user wants transparency, use a two-step process: generate with a chroma key background, then remove it with the image-tools:manipulate-image skill afterwards.

Step 1 — Analyze subject colors BEFORE choosing chroma key

MANDATORY: Before picking a chroma key color, list every color the subject will contain. Think through:

• Skin/body color (e.g., grey hippo, brown dog, pink flamingo)
• Clothing colors (shirt, pants, shoes, accessories)
• Hair/bow/hat colors
• Object colors (bike, tools, props)
• Style colors (cartoon outlines are typically black)

Then pick the chroma key that has zero overlap with any of those colors.

Step 2 — Pick from the chroma key palette

Chroma Key	Exact RGB	Hex	Gemini keyword	Use when subject does NOT contain
Green	RGB(0, 255, 0)	#00FF00	"chroma key green"	Green, lime, emerald, forest tones
Magenta	RGB(255, 0, 255)	#FF00FF	"chroma key magenta"	Pink, purple, magenta, violet, fuchsia tones
Blue	RGB(0, 0, 255)	#0000FF	"chroma key blue"	Blue, sky blue, navy, cyan, teal tones
Yellow	RGB(255, 255, 0)	#FFFF00	"chroma key yellow"	Yellow, gold, blonde, sunshine, amber tones

Selection rules:

1. List ALL expected subject colors first
2. Eliminate any chroma key that overlaps with subject colors
3. From remaining options, pick the one with maximum hue distance from the subject's dominant color
4. Green is the safest default — most cartoon/illustration subjects don't contain bright lime green
5. Never use magenta for subjects with pink, purple, or magenta clothing/accessories
6. Never use blue for sky scenes, water, or subjects wearing blue
7. Never use yellow for warm-lit scenes, blonde hair, or gold elements
8. Document your color analysis in the --composition flag

Step 3 — Generate with chroma key prompt

Append this exact block to the end of your prompt (replace [COLOR], [R,G,B], [HEX]):

The subject is placed on a chroma key [COLOR] studio backdrop (RGB [R,G,B], hex [HEX]). Plain studio setting with soft, diffused lighting on the subject only. The background is a perfectly flat, uniform solid [COLOR] with no gradients, no shadows, no texture, no patterns, no depth of field, no bokeh. Every pixel of the background must be the same [COLOR] color (RGB [R,G,B], #[HEX]) from edge to edge with zero variation.

Example for green chroma key:

The subject is placed on a chroma key green studio backdrop (RGB 0,255,0, hex #00FF00). Plain studio setting with soft, diffused lighting on the subject only. The background is a perfectly flat, uniform solid green with no gradients, no shadows, no texture, no patterns, no depth of field, no bokeh. Every pixel of the background must be the same green color (RGB 0,255,0, #00FF00) from edge to edge with zero variation.

Step 4 — Remove background (post-processing)

Gemini won't produce exact RGB values, so always sample the actual corner pixel, then use the manipulate-image scripts directly to remove the background and trim.

VENV="${CLAUDE_PLUGIN_ROOT}/scripts/venv"
MANIPULATE="${CLAUDE_PLUGIN_ROOT}/skills/manipulate-image/scripts"

# 1. Sample actual background color from corner
"$VENV/bin/python" -c "from PIL import Image; print(Image.open('<image>').getpixel((0,0)))"

# 2. Remove chroma key with HSV detection + spill suppression
"$MANIPULATE/run.sh" alpha <image> --transparent "<actual R,G,B>" --tolerance 15 --feather 40 -o <output>

# 3. Trim empty space
"$MANIPULATE/run.sh" trim <output> -o <final>

When to use this

User mentions "transparent", "no background", "PNG with alpha", "cutout", or "sticker". Always tell the user you're using a chroma key approach since the result won't be transparent until the second step.

Reference

For detailed setup, troubleshooting, prompt engineering, and file templates, see SETUP.md.