Imagen on Vertex AI model versions and lifecycle

Each Vertex AI Generative AI image model is available in distinct versions. The image models include generation and text models, such as imagegeneration and imagetext. The versions differ by the features they offer.

Image generation and editing

The following image generation model versions are available for image generation and image editing tasks.

Imagen 3 Imagen 2 (v.006, default) Imagen 2 (v.005) Imagen (v.002)
Model name and version Imagen 3: imagen-3.0-generate-001

Imagen 3 Fast: imagen-3.0-fast-generate-001

This is a low-latency model variant you can use for prototyping or low-latency use cases.

Imagen 3 Customization and Editing: imagen-3.0-capability-001 		imagegeneration@006 imagegeneration@005 imagegeneration@002
Image generation General Availability General Availability (approved users) General Availability (approved users) General Availability (approved users)
Image Customization (few-shot learning) General Availability (approved users)
  • Subject customization (product, person, and animal companion)
  • Style customization
  • Controlled customization (scribble and canny edge)
  • Instruct customization (style transfer)
 Not supported Not supported Not supported
Image editing
  • Editing (mask-based)
    • Inpainting (insert, remove)
    • Outpainting
    • Product image editing
  • Editing (mask-based)
    • Inpainting (insert, remove)
    • Outpainting
    • Product image editing
 Not supported. Feature support for this model is subject to future changes and updates.
  • Editing (mask-based)
  • Editing (mask-free)
  • Product image editing (Preview feature)
Digital watermarking and verification (image generation) Supported (enabled by default) Supported (enabled by default) Supported (not enabled by default) Supported (not enabled by default)
User-configurable safety settings
  • Person generation - allow all ages, only adult person/face generation, or block all person generation
  • Safety setting - block most, some, or few instances of filtered content
  • Person generation - allow all ages, only adult person/face generation, or block all person generation
  • Safety setting - block most, some, or few instances of filtered content
 Not supported Not supported
Image upscaling Not supported Not supported Not supported Supported (2x or 4x)
Aspect ratio and resolution
  • 1:1 - 1024x1024 pixels (square)
  • 9:16 - 768x1408 pixels
  • 16:9 - 1408x768 pixels
  • 3:4 - 896x1280 pixel
  • 4:3 - 1280x896 pixels
  • 1:1 - 1536x1536 pixels (square)
  • 9:16 - 1152x2016 pixels
  • 16:9 - 2016x1134 pixels
  • 3:4- 1344x1792 pixel
  • 4:3 - 1792x1344 pixels
  • 1:1 - 1536x1536 pixels (square)
  • 9:16 - 1152x2016 pixels
  • 1:1 - 1024x1024 pixels (square)
Languages

General availability:

  • English

Preview:

  • Chinese (simplified)
  • Chinese (traditional)
  • Hindi
  • Japanese
  • Korean
  • Portuguese
  • Spanish

General availability:

  • English

Preview:

  • Chinese (simplified)
  • Chinese (traditional)
  • Hindi
  • Japanese
  • Korean
  • Portuguese
  • Spanish

General availability:

  • English

Preview:

  • Chinese (simplified)
  • Chinese (traditional)
  • Hindi
  • Japanese
  • Korean
  • Portuguese
  • Spanish

General availability:

  • English
Model tuning Not supported.

Feature support for this model is subject to future changes and updates.		 Not supported Not supported

Model tuning availability:

  • Subject model tuning
  • Style model tuning

Visual captioning and Visual Question Answering (VQA)

The following imagetext model versions are available for image captioning and Visual Question Answering (VQA) tasks.

Model version:
imagetext@001 (default)
Visual captioning (short-form) Tokens: 32 (~32 words)
Languages: English, Spanish, French, German, Italian
Visual Question Answering, VQA (short-form) Tokens (request): 80
Tokens (response): 64
Languages: English

How to specify the version of an image model

To use a version of a vision model, add the three digit version number to the model. For example, to specify an imagegeneration model version, append the model version (for example, @006) to the model's name:

  • POST https://us-central1-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/us-central1/publishers/google/models/imagegeneration@006:predict
  • POST https://us-central1-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/us-central1/publishers/google/models/imagetext@001:predict

However, this format doesn't apply to Imagen 3 models.

You should always specify a stable model version for production use, as using the default option without a model version specified can lead to changes in model behavior.

If you omit the model version, Imagen uses the default model version:

  • POST https://us-central1-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/us-central1/publishers/google/models/imagegeneration:predict
  • POST https://us-central1-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/us-central1/publishers/google/models/imagetext:predict

As a best practice, always specify a model version. If you choose to use the model name without a model version, make sure it is in a testing environment.

What's next