Official

google / imagen-3-fast

A faster and cheaper Imagen 3 model, for when price or speed are more important than final image quality

  • Public
  • 30K runs
  • Paper
Run with an API
Playground API Examples README

Examples

View more

Pricing

Official model
Pricing for official models works differently from other models. Instead of being billed by time, you’re billed by input and output, making pricing more predictable.
Learn more

This model is priced by how many images are generated.

Check out our docs for more information about how per-image pricing works on Replicate.

Readme

Imagen 3

Imagen 3 is DeepMind’s latest text-to-image generative model, focusing on high-quality image generation with improved detail, lighting, and reduced artifacts.

Core Capabilities

  • Enhanced prompt understanding for complex image generation tasks
  • Improved text rendering for applications like presentations and typography
  • Support for diverse artistic styles from photorealism to animation
  • Better handling of lighting, textures, and fine details
  • Natural language prompt processing without requiring complex prompt engineering

Technical Improvements

Image Quality

  • Enhanced color balance and vibrancy
  • Improved texture rendering
  • Better detail preservation in complex scenes
  • Reduced artifact generation
  • More accurate style reproduction across different artistic genres

Prompt Processing

  • Support for longer, more detailed prompts
  • Better understanding of camera angles and composition requirements
  • Improved handling of specific style requests
  • Enhanced text rendering capabilities

Benchmarks

Performance metrics based on human evaluation using GenAI-Bench:

  • Highest score for visual quality among compared models
  • High accuracy in prompt response adherence
  • Strong performance in overall preference benchmarks

Detailed benchmark methodology and results are available in Appendix D of the technical report.

Security Features

  • Built-in content filtering system
  • Dataset filtering to minimize harmful content
  • SynthID watermarking integration for image identification
  • Extensive red teaming and evaluations for: Fairness, Bias, Content safety

Technical Documentation

For detailed technical specifications and methodology, refer to the full technical report.

Integration

SynthID watermarking is integrated by default, embedding digital watermarks directly into image pixels while remaining visually imperceptible.

Development Team

Core development involved collaboration across multiple technical disciplines including:

  • Machine learning research
  • Computer vision
  • Natural language processing
  • Security engineering
  • Dataset engineering

For a complete list of contributors and their roles, refer to the technical report.

Privacy

Data from this model is sent from Replicate to Google.

Check their Privacy Policy for details:

https://policies.google.com/privacy