"""generate_image_nano — Google Imagen 3 fast text-to-image generation.
"Nano Banana Pro" — fast, cheap concept generation using Google's Imagen 3
via Vertex AI. Great for rapid iteration before committing to SD 3.5L renders.
Cost: 1 GCX ($0.10) vs SD 3.5L at 2 GCX ($0.20).
"""
import base64
import io
import structlog
from google.cloud import aiplatform
from vertexai.preview.vision_models import ImageGenerationModel
from ..mcp_server.config import config
logger = structlog.get_logger()
_model = None
def _get_model() -> ImageGenerationModel:
global _model
if _model is None:
aiplatform.init(
project=config.gcp_project,
location=config.gcp_region,
)
_model = ImageGenerationModel.from_pretrained("imagen-3.0-fast-generate-001")
return _model
def generate_image_nano(
prompt: str,
width: int = 1024,
height: int = 1024,
enhance_prompt: bool = True,
) -> dict:
"""Generate an image using Google Imagen 3 (fast variant).
Args:
prompt: Text description of the desired image.
width: Output width (supported: 1024, 2048).
height: Output height (supported: 1024, 2048).
enhance_prompt: Let Imagen enhance your prompt for better results.
Returns:
dict with image (base64 PNG), width, height, model info.
"""
# Imagen 3 supports specific aspect ratios, normalize to closest
# Supported: 1:1 (1024x1024), 3:4, 4:3, 9:16, 16:9
if width > 1536 or height > 1536:
width = 2048 if width >= height else 1024
height = 2048 if height >= width else 1024
else:
width = 1024
height = 1024
logger.info("generate_image_nano", prompt=prompt[:80], width=width, height=height)
# Content safety: run the same prompt enhancer if available
original_prompt = prompt
enhancement_info = None
if enhance_prompt:
try:
from .prompt_enhancer import enhance_prompt as _enhance
enhancement = _enhance(prompt)
enhancement_info = enhancement
if enhancement.get("status") == "rejected":
raise ValueError(
f"Content policy violation: {enhancement.get('rejection_reason', 'Prompt rejected')}"
)
prompt = enhancement.get("enhanced_prompt", prompt)
logger.info("generate_nano.enhanced", original=original_prompt[:60], enhanced=prompt[:60])
except ValueError:
raise
except Exception as e:
logger.warning("generate_nano.enhance_failed", error=str(e))
model = _get_model()
response = model.generate_images(
prompt=prompt,
number_of_images=1,
add_watermark=False, # We handle provenance via Golden Codex
)
if not response.images:
raise RuntimeError("Imagen 3 returned no images")
# Convert to base64
image = response.images[0]
buffer = io.BytesIO()
image._pil_image.save(buffer, format="PNG")
image_b64 = base64.b64encode(buffer.getvalue()).decode("utf-8")
result = {
"image": image_b64,
"width": width,
"height": height,
"model": "google/imagen-3.0-fast",
"prompt": prompt,
}
if enhancement_info and enhancement_info.get("status") == "approved":
result["original_prompt"] = original_prompt
result["enhanced_prompt"] = prompt
result["prompt_changes"] = enhancement_info.get("changes_summary", "")
return result
