Gemini Image Gen

Gemini Image Generation

Generate images via google/gemini-3.1-flash-image-preview on OpenRouter. Cheap ($0.25/M in, $1.5/M out), fast, good quality.

Quick Start

python3 scripts/generate.py "a watercolor illustration of a cozy café" -o output.png

With reference image (style/character guidance):

python3 scripts/generate.py "same character but waving hello" -o wave.png --ref reference.png

Script path: skills/gemini-image-gen/scripts/generate.py

Requirements

• OPENROUTER_API_KEY environment variable (or --api-key flag)
• Python 3.10+ (stdlib only, no pip installs needed)

How It Works

1. Calls OpenRouter /chat/completions with modalities: ["text", "image"]
2. Optionally encodes a reference image as base64 in the message
3. Extracts generated image from choices[0].message.images[0].image_url.url (data:image/png;base64,...)
4. Decodes and saves to output path

Prompt Engineering Tips (from experience)

Aspect Ratio & Composition

• Gemini respects aspect ratio instructions in the prompt
• For vertical (e.g. phone wallpaper, Xiaohongshu cover): add "vertical composition, 3:4 aspect ratio"
• For horizontal (e.g. banner): add "horizontal composition, 16:9 aspect ratio"
• For square: add "square composition, 1:1 aspect ratio"
• Always specify — without it, Gemini defaults to roughly square and may crop awkwardly

Character Consistency

• When using --ref, describe the character features explicitly in the prompt AND provide the reference image
• Key details to specify: hair color/style, eye color, clothing, accessories, expression
• Example: "same character from reference: silver-to-ice-blue gradient shoulder-length hair, ice-blue eyes, cream cardigan over light blue shirt, snowflake earring"
• Gemini is decent at maintaining consistency but drifts on small details — always re-specify distinguishing features

Style Control

• Name the art style explicitly: "soft watercolor illustration", "anime cel-shading", "photorealistic", "flat vector", "oil painting"
• For warm/cozy tone: "warm color palette, cream and peach gradient background, bokeh light spots"
• For dark/moody: "dark gradient background, deep navy to black, subtle glow effects"
• Mentioning a well-known art style works: "in the style of Studio Ghibli", "Makoto Shinkai lighting"

Text in Images

• Gemini can render short text in images but it's unreliable for CJK characters
• For English text: works reasonably well if you specify font style ("bold sans-serif", "handwritten script")
• For Chinese/Japanese: avoid — it usually garbles characters. Add text overlays with a separate tool (e.g. ImageMagick, Pillow) instead

Common Pitfalls

• Body proportions: Gemini sometimes compresses/distorts figures. Add "natural human body proportions, do not squash or stretch" for character art
• Hands: Still a weak spot. Minimize visible hands or describe hand pose explicitly
• Multiple subjects: More than 2-3 subjects increases inconsistency. Keep scenes focused
• Batch generation: For generating multiple variations, run the script multiple times — each call is independent. Do NOT ask for "4 options" in one prompt

Sending Images on Feishu

⚠️ Critical: Images must be saved to a path within localRoots (typically your OpenClaw workspace dir). /tmp is NOT whitelisted on Feishu.

# Save to workspace, not /tmp
output_path = "my_image.png"  # relative to workspace

# Send via message tool:
#   media: "file://<workspace_path>/my_image.png"
#   (use 'media' parameter, NOT 'filePath')

After sending, clean up temporary images to avoid workspace clutter.

Advanced: Calling from Python (without CLI)

import os, sys
sys.path.insert(0, "skills/gemini-image-gen/scripts")
from generate import generate

generate(
    prompt="a cute robot reading a philosophy book",
    output="robot.png",
    ref_image=None,  # or path to reference image
)

Model Alternatives

Troubleshooting

• "No image in response": Check .debug.json file created alongside output. Usually means the prompt triggered safety filters or the model returned text-only.
• Garbled/distorted output: Try rephrasing. Add "high quality, detailed" and be more specific about composition.
• API error 429: Rate limited. Wait 30s and retry.
• API error 402: Insufficient credits on OpenRouter.