Generative AI Google Brain

gemini-use-cases-media-generation-consistent-imagery-generation

3 min read

Generating Consistent Imagery with Gemini 🍌

Source notebook

Repo path: gemini/use-cases/media-generation/consistent_imagery_generation.ipynb · Open on GitHub · intermediate

Generates a consistent robot image story from one archive image using Gemini 2.5 Flash Image.

Summary

This notebook teaches how to reuse an existing image as a visual reference and generate a sequence of consistent new images with Gemini 2.5 Flash Image. It sets up the Google Gen AI SDK, loads an archive image, creates a robot character sheet, generates successive scenes from prior assets, visualizes the asset lineage as a directed graph, and saves prompts plus source links in PNG metadata.

Key code patterns

Unified Gen AI client

from google import genai
 
check_environment()
client = genai.Client()
check_configuration(client)

Creates one SDK client that can use either Agent Platform or Google AI Studio configuration.

Image generation config

GENERATION_CONFIG = GenerateContentConfig(
    response_modalities=["IMAGE"],
    image_config=ImageConfig(aspect_ratio="16:9"),
)

Constrains output to image responses and requests a supported 16:9 aspect ratio.

Reference-guided generation

contents = [*sources, prompt] if sources else prompt
response = client.models.generate_content(
    model="gemini-2.5-flash-image",
    contents=contents,
    config=GENERATION_CONFIG,
)

Combines prior PIL images with a prompt so new images preserve character and scene context.

Retry transient client errors

tenacity.Retrying(
    stop=tenacity.stop_after_attempt(7),
    wait=tenacity.wait_incrementing(start=10, increment=1),
    retry=tenacity.retry_if_exception(should_retry_request),
    reraise=True,
)

Retries selected 400 and 429 ClientError cases called out for first Cloud Storage access and low quota.

Asset lineage graph

graph = nx.DiGraph(assets=assets)
for asset in assets.values():
    graph.add_node(asset.id, asset=asset)
    for source_id in asset.source_ids:
        graph.add_edge(source_id, asset.id)

Tracks generated-from relationships between archive, character sheet, and scene assets.

Prompt metadata persistence

metadata = PngInfo()
metadata.add_text("source_ids", "\n".join(asset.source_ids))
metadata.add_text("prompt", asset.prompt)
asset.pil_image.save(image_path, pnginfo=metadata)

Stores source asset IDs and prompts inside PNG metadata for local restore.

Models & APIs used

  • Models: gemini-2.5-flash-image
  • APIs / services: Agent Platform, Gemini API, Google AI Studio
  • SDKs / libraries: google-genai, networkx, tenacity, matplotlib, pillow

When to use this

Use this pattern to generate a consistent visual story or asset set from existing images using reference images and iterative prompts.

Gotchas & caveats

  • Agent Platform requires a Google Cloud project and the Agent Platform API enabled.
  • Preview models must use location global.
  • Google AI Studio requires a Gemini API key.
  • Colab secrets or environment variables are needed for SDK configuration outside managed environments.
  • Temporary projects may hit 1 QPM quota and trigger 429 retries.
  • First Cloud Storage access in a workshop project may trigger a retryable 400 while service agents are provisioned.
  • Repeated generations can vary when prompt details are left open.
  • Closed eyes were described as difficult to get consistently after many accumulated transformations.

Best practices

  • Use environment variables or Colab Secrets instead of hardcoding API configuration.
  • Use a character sheet as a reusable design reference for future image-generation tasks.
  • Refer explicitly to input images, such as Image 1 and Image 2, to avoid ambiguity.
  • Clarify removed or changed objects, such as no longer holding the map or removing ice axes.
  • Spend time refining the first scene because it cascades into later generated scenes.
  • Use previous scenes plus the character sheet to preserve context across successive generations.
  • Save prompts and source IDs in image metadata so assets can be restored without a database.
  • Focus significant changes over iterative steps or create character sheets upfront for more deterministic control.

Graph View

Backlinks