razbakov/image-from-gemini

Gemini Image Generation

Generate images from text prompts using Google's Gemini API, optionally with uploaded reference photos.

Requirements

• GEMINI_API_KEY environment variable (or set in ~/.zshrc)
• python3 (standard on macOS/Linux)
• Get a key at https://aistudio.google.com/apikey

Workflow

1. Text-only generation

python3 ~/.claude/skills/image-from-gemini/scripts/generate.py \
  "your detailed image prompt here" \
  -o /path/to/output.png

2. Photo-to-poster generation

Upload one or more reference photos. Gemini will incorporate the people into the generated design.

# Single photo
python3 ~/.claude/skills/image-from-gemini/scripts/generate.py \
  -i /path/to/person.jpg \
  "Square Instagram post. Place the person from the photo as the hero..." \
  -o /path/to/output.png

# Multiple photos
python3 ~/.claude/skills/image-from-gemini/scripts/generate.py \
  -i /path/to/person1.jpg \
  -i /path/to/person2.jpg \
  -i /path/to/person3.jpg \
  "Instagram story with all uploaded people as instructors..." \
  -o /path/to/output.png

3. Display and iterate

After generation, use the Read tool to display the image inline. Adjust the prompt and regenerate as needed.

4. Batch generation (multiple styles)

For generating multiple variants, use Python directly to avoid repeated shell calls:

python3 << 'EOF'
import base64, json, os, urllib.request

api_key = os.environ.get("GEMINI_API_KEY")
model = "gemini-3.1-flash-image-preview"
url = f"https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent?key={api_key}"

with open("photo.jpg", "rb") as f:
    img_b64 = base64.b64encode(f.read()).decode()

prompts = {
    "v1-style": "prompt for style 1...",
    "v2-style": "prompt for style 2...",
}

for name, prompt in prompts.items():
    payload = {
        "contents": [{"parts": [
            {"inlineData": {"mimeType": "image/jpeg", "data": img_b64}},
            {"text": prompt}
        ]}],
        "generationConfig": {"responseModalities": ["TEXT", "IMAGE"]}
    }
    data = json.dumps(payload).encode()
    req = urllib.request.Request(url, data=data, headers={"Content-Type": "application/json"})
    resp = urllib.request.urlopen(req, timeout=180)
    result = json.loads(resp.read())
    for candidate in result.get("candidates", []):
        for part in candidate.get("content", {}).get("parts", []):
            if "inlineData" in part:
                img_data = base64.b64decode(part["inlineData"]["data"])
                with open(f"output-{name}.png", "wb") as f:
                    f.write(img_data)
                break
EOF

Photo-to-poster prompting tips

When uploading photos of real people for event posters:

• Say "Place the person from the photo" — Gemini understands the reference
• Say "remove the original background" — Gemini will cut out the person
• Use rembg (Python) beforehand for cleaner cutouts if needed: uvx --with "rembg[cpu,cli]" rembg i input.jpg output.png
• Specify exact text: spell out every word, every number, every address
• Repeat critical text in the prompt (addresses, URLs) to avoid hallucination
• Say "DO NOT include any logo" if you plan to overlay it later with HTML
• For multiple photos, upload all and number them: "Photo 1: Name1, Photo 2: Name2"

Style presets

Proven styles for event/dance posters:

| Style | Description | Best for | |-------|------------|----------| | Neon | Dark bg, neon pink/cyan streaks, nightclub vibe | Club events, parties | | Cuban | Cuban flag colors, distressed texture, palm shadows, Havana aesthetic | Latin dance, cultural events | | Minimal | Black bg, gold circle frame, Swiss typography | Premium/luxury feel | | Fire | Flames, sparks, dramatic lighting from below | High-energy concerts, competitions | | Warm bokeh | Burgundy/maroon, golden bokeh lights | Warm, inviting classes |

General prompting tips

• Be specific about layout, colors, fonts, and mood
• For posters with text: spell out every word exactly as it should appear
• Mention dimensions: "Square 1080x1080" or "Portrait 1080x1920"
• Gemini avoids contractions — use "will not" instead of "won't"
• For dark backgrounds, specify hex colors (e.g., "#1E1B2E")

Available models

| Model | Best for | |-------|----------| | gemini-3.1-flash-image-preview | Fast, good quality (default) | | gemini-3-pro-image-preview | Higher quality, slower |

Troubleshooting

• "API key not found" — Set export GEMINI_API_KEY=your-key in ~/.zshrc
• 404 model not found — List models: curl -s "https://generativelanguage.googleapis.com/v1beta/models?key=$GEMINI_API_KEY" | python3 -c "import json,sys; [print(m['name']) for m in json.load(sys.stdin)['models'] if 'image' in m['name']]"
• Argument list too long — The script uses urllib internally, not curl. If using the API directly, avoid subprocess with large base64 payloads.
• 500 Internal Server Error — Image may be too small or corrupted. Use the original full-resolution photo, not a tiny crop.
• No image returned — Rephrase or simplify the prompt. Gemini may refuse certain compositions.