Imagen Generation Skill

When to Invoke

Skill({ skill: 'imagen-generation' });

Use when:

• Generating images from text prompts via Google Imagen on Vertex AI
• Editing existing images with text instructions (image-to-image)
• Inpainting / outpainting specific regions of an image
• Upscaling images using Imagen upscaler
• Integrating AI image generation into Python workflows

---

Setup

Prerequisites

pip install google-cloud-aiplatform pillow
gcloud auth application-default login
gcloud config set project YOUR_PROJECT_ID

Enable Vertex AI

gcloud services enable aiplatform.googleapis.com

Authentication

import vertexai
from vertexai.preview.vision_models import ImageGenerationModel

vertexai.init(project='YOUR_PROJECT_ID', location='us-central1')

---

Text-to-Image Generation

import vertexai
from vertexai.preview.vision_models import ImageGenerationModel
from PIL import Image
import io

vertexai.init(project='YOUR_PROJECT_ID', location='us-central1')

model = ImageGenerationModel.from_pretrained('imagegeneration@006')

# Basic generation
response = model.generate_images(
    prompt='A futuristic city skyline at sunset, photorealistic, 4K',
    number_of_images=1,
    aspect_ratio='1:1',        # '1:1', '9:16', '16:9', '3:4', '4:3'
    guidance_scale=7.5,        # 1-20; higher = closer to prompt
    seed=42,                   # Optional: for reproducibility
)

# Save the image
image = response.images[0]
image.save('output.png')

# Or convert to PIL
pil_image = Image.open(io.BytesIO(image._image_bytes))
pil_image.show()

---

Batch Generation

prompts = [
    'A serene mountain lake at dawn',
    'Abstract digital art with geometric shapes',
    'A cozy coffee shop interior',
]

for i, prompt in enumerate(prompts):
    response = model.generate_images(
        prompt=prompt,
        number_of_images=1,
    )
    response.images[0].save(f'image_{i}.png')
    print(f'Saved image_{i}.png for: {prompt[:50]}')

---

Image Editing (Image-to-Image)

from vertexai.preview.vision_models import ImageGenerationModel, Image as VertexImage

model = ImageGenerationModel.from_pretrained('imagegeneration@006')

# Load source image
source_image = VertexImage.load_from_file('source.png')

response = model.edit_image(
    base_image=source_image,
    prompt='Make the sky more dramatic with storm clouds',
    edit_mode='inpainting-insert',   # 'inpainting-insert' | 'inpainting-remove' | 'outpainting'
    mask_mode='background',           # 'background' | 'foreground' | 'semantic'
    number_of_images=1,
    guidance_scale=8.0,
)

response.images[0].save('edited.png')

---

Inpainting with Mask

import numpy as np
from PIL import Image, ImageDraw

# Create a mask (white = area to inpaint, black = keep)
source_pil = Image.open('source.png')
mask = Image.new('L', source_pil.size, 0)  # Black background
draw = ImageDraw.Draw(mask)
draw.rectangle([100, 100, 300, 300], fill=255)  # White region to replace
mask.save('mask.png')

# Load for Vertex AI
source_image = VertexImage.load_from_file('source.png')
mask_image = VertexImage.load_from_file('mask.png')

response = model.edit_image(
    base_image=source_image,
    mask=mask_image,
    prompt='A beautiful garden fountain',
    edit_mode='inpainting-insert',
    number_of_images=1,
)
response.images[0].save('inpainted.png')

---

Upscaling

from vertexai.preview.vision_models import ImageGenerationModel, Image as VertexImage

model = ImageGenerationModel.from_pretrained('imagegeneration@006')
source_image = VertexImage.load_from_file('low_res.png')

response = model.upscale_image(
    image=source_image,
    upscale_factor='x2',   # 'x2' or 'x4'
)
response.save('upscaled.png')

---

Negative Prompts

Use negative prompts to exclude unwanted elements:

response = model.generate_images(
    prompt='Portrait of a professional business person in an office',
    negative_prompt='blurry, low quality, cartoon, anime, watermark, text, logo',
    number_of_images=2,
    guidance_scale=9.0,
)
for i, img in enumerate(response.images):
    img.save(f'portrait_{i}.png')

---

Imagen 3 (Latest Model)

# Imagen 3 — highest quality, best prompt adherence
model = ImageGenerationModel.from_pretrained('imagen-3.0-generate-001')

response = model.generate_images(
    prompt='A photorealistic macro photograph of a dewdrop on a spider web at sunrise',
    number_of_images=1,
    aspect_ratio='3:4',
    safety_filter_level='block_some',  # 'block_most' | 'block_some' | 'block_few'
    person_generation='allow_adult',   # 'dont_allow' | 'allow_adult'
)
response.images[0].save('imagen3_output.png')

---

Model Reference

Model ID	Use Case	Notes
imagen-3.0-generate-001	Highest quality generation	Latest, best prompt follow
imagen-3.0-fast-generate-001	Fast/cost-effective generation	Lower latency
imagegeneration@006	Stable production model	Well-tested
imagegeneration@005	Previous generation	Legacy
imagen-3.0-capability-001	Editing and transformations	Inpaint, outpaint

---

Hugging Face Alternative (No GCP Required)

For local or non-GCP environments, use Stable Diffusion via diffusers:

pip install diffusers transformers accelerate torch

from diffusers import StableDiffusionPipeline
import torch

pipe = StableDiffusionPipeline.from_pretrained(
    'runwayml/stable-diffusion-v1-5',
    torch_dtype=torch.float16,
)
pipe = pipe.to('cuda')  # or 'cpu' (slow)

image = pipe(
    prompt='A futuristic city at sunset',
    negative_prompt='blurry, low quality',
    num_inference_steps=30,
    guidance_scale=7.5,
).images[0]

image.save('output.png')

---

Cost Optimization

• Use imagen-3.0-fast-generate-001 for iteration/drafts; switch to imagen-3.0-generate-001 for final output
• Generate 1-2 images per call during development; batch only in production
• Cache results when the same prompt is used repeatedly
• Use seed for reproducibility to avoid regenerating identical images
• Vertex AI pricing: check cloud.google.com/vertex-ai/pricing (billed per image)

---

Safety and Content Policy

• Imagen enforces Google's content policy; explicit/harmful content requests are blocked
• safety_filter_level controls strictness: block_most (safest) → block_few (permissive)
• person_generation='dont_allow' disables human face generation for child-safety compliance
• Store generation_parameters from response for audit/reproducibility requirements
• Never generate images of real people without appropriate consent handling