polym-eval-generate-gemini

name	polym-eval-generate-gemini
description	Generate Gemini/Imagen images via Vertex AI, AI Studio, or AIDP. Use for Gemini image generation, editing, and reference-image workflows.

Generate Gemini Images

Generate images using Google Gemini (nano banana / nano banana 2) via two backends:

• google (default): Vertex AI + Google AI Studio with automatic fallback
• aidp: ByteDance AIDP model hub (internal, fastest for ByteDance users)

Two Backends

Backend	Models	Endpoint	Auth
google	gemini-2.0-flash-preview-image-generation	Vertex AI / AI Studio	GOOGLE_AI_STUDIO_API_KEY
aidp	gemini-3.1-fi (Nano Banana 2, default), gemini-2.5-flash-image (Nano Banana), gemini_nbp (Nano Banana Pro (Gemini 3.0 Image))	aidp.bytedance.net/.../multimodal/crawl	AIDP_API_KEY

Pre-flight Check (IMPORTANT - Do this BEFORE generating)

Before running any generation command, the Agent MUST check that required environment variables are configured:

1. Read the .env file at the project root and verify:

   • google backend: GOOGLE_AI_STUDIO_API_KEY (required), optional: GOOGLE_CLOUD_PROJECT, GOOGLE_CLOUD_LOCATION, GEMINI_IMAGE_API_KEY
   • aidp backend: AIDP_API_KEY

2. If any required variable is missing, tell the user which variables are needed and ask for values before proceeding.

Quick Start

# Google backend (default)
python skills/polym-eval-generate-gemini/scripts/generate.py --prompt "A serene mountain landscape at sunset"

# AIDP backend (ByteDance internal)
python skills/polym-eval-generate-gemini/scripts/generate.py --prompt "A serene mountain landscape at sunset" --backend aidp

Required Configuration

Set in your .env file:

# Google backend
GOOGLE_CLOUD_PROJECT=your-project-id
GOOGLE_CLOUD_LOCATION=global
GEMINI_IMAGE_API_KEY=your_vertex_api_key  # or use ADC
GOOGLE_AI_STUDIO_API_KEY=your_ai_studio_key

# AIDP backend
AIDP_API_KEY=your_aidp_key

# Optional: Model overrides
GEMINI_IMAGE_MODEL=gemini-2.0-flash-preview-image-generation  # google backend default
GEMINI_AIDP_MODEL=gemini-3.1-fi                               # aidp backend default

Usage Examples

AIDP: Basic text-to-image

python skills/polym-eval-generate-gemini/scripts/generate.py \
  --prompt "A photorealistic cat sitting on a windowsill" \
  --backend aidp \
  --output generated/cat.png

AIDP: Image editing with reference

python skills/polym-eval-generate-gemini/scripts/generate.py \
  --prompt "Make this image look like a watercolor painting" \
  --backend aidp \
  --ref-images path/to/photo.jpg \
  --output generated/watercolor.png

AIDP: Use nano banana pro model

python skills/polym-eval-generate-gemini/scripts/generate.py \
  --prompt "A product photo on white background" \
  --backend aidp \
  --model gemini_nbp \
  --output generated/product.png

Google: Basic text-to-image

python skills/polym-eval-generate-gemini/scripts/generate.py \
  --prompt "A photorealistic cat sitting on a windowsill" \
  --output generated/cat.png

Google: Specify size

python skills/polym-eval-generate-gemini/scripts/generate.py \
  --prompt "Abstract geometric art" \
  --size 2K \
  --output generated/abstract.png

Google: Force AI Studio instead of Vertex AI

python skills/polym-eval-generate-gemini/scripts/generate.py \
  --prompt "A sunset over the ocean" \
  --use-ai-studio

Complete Parameter Reference

Parameter	CLI Flag	Type	Description	Default
prompt	--prompt, -p	string	Text description	Required
backend	--backend, -b	string	google or aidp	google
model	--model, -m	string	Model ID	see below
size	--size, -s	string	1K, 2K, 4K (google only)	1K
ref_images	--ref-images, -r	list	Reference image paths/URLs	None
use_ai_studio	--use-ai-studio	flag	Force AI Studio (google only)	False
no_retry	--no-retry	flag	Don't fallback to AI Studio (google only)	False
output	--output, -o	string	Output file path	generated/{timestamp}.png

Available Models

AIDP Backend

AIDP Model ID	Google Model Name	Alias	Speed	Notes
gemini-3.1-fi	Gemini 3.1 Flash Image Preview	Nano Banana 2 (Gemini 3.1 Flash Image Preview)	~35s	Default; has internal thinking
gemini-2.5-flash-image	Gemini 2.5 Flash Image	Nano Banana (Gemini 2.5 Flash Image)	~8s	Faster, no thinking
gemini_nbp	Gemini 3.0 Image	Nano Banana Pro (Gemini 3.0 Image)	unknown	Previous generation

Google Backend

Model ID	Notes
gemini-2.0-flash-preview-image-generation	Default

AIDP API Details

• Endpoint: https://aidp.bytedance.net/api/modelhub/online/multimodal/crawl
• Auth header: api-key: <AIDP_API_KEY>
• Request format: OpenAI chat completions JSON
• Response: choices[0].message.multimodal_contents — array with items of type inline_data containing base64-encoded image

Response image extraction:

multimodal_contents = response["choices"][0]["message"]["multimodal_contents"]
image_item = next(i["inline_data"] for i in multimodal_contents if i["type"] == "inline_data")
image_bytes = base64.b64decode(image_item["data"])

Programmatic Usage

# Google backend
from generate import generate_image

result = generate_image(
    id="my_image",
    prompt="A beautiful sunset over the ocean",
    ref_image_paths=[],
    path="generated",
    size="1K",
    model="gemini-2.0-flash-preview-image-generation",
)

# AIDP backend
from generate import aidp_generate_image

result = aidp_generate_image(
    prompt="A beautiful sunset over the ocean",
    ref_image_paths=[],
    output_file="generated/output.png",
    model="gemini-3.1-fi",
)

# Returns: {"gen_success": bool, "gen_image_path": str, "gen_url": str}

Dependencies

google-genai
Pillow
requests
python-dotenv

Install:

pip install google-genai Pillow requests python-dotenv

Troubleshooting

"No image in response"

The model may have filtered the prompt. Try rephrasing or check prompt_feedback in logs.

429 Rate Limit (Google backend)

Configure both Vertex AI and AI Studio keys for automatic fallback.

AIDP timeout

gemini-3.1-fi (default) takes 30–60 seconds due to internal thinking/reasoning. Use gemini-2.5-flash-image for faster generation (~8s) without thinking.

"product[gemini_3.1_fi] request should be process by cluster[multimodal...]"

Using wrong endpoint — must use multimodal/crawl, not v2/crawl.