Creativity · AI Model

Imagen 3

Imagen 3 is Google DeepMind's text-to-image model, generally available from late 2024 on Vertex AI and the Gemini API. It produces high-fidelity photorealistic images, renders text inside images more reliably than prior diffusion models, and watermarks every output with SynthID for provenance.

Model specs

Vendor
Google
Family
Imagen
Released
2024-12
Context window
1,024 tokens
Modalities
text, vision
Input price
n/a
Output price
n/a
Pricing as of
2026-04-20

Strengths

  • Strong photorealism and fine-detail rendering
  • Better in-image typography than most diffusion rivals
  • Every image tagged with SynthID for provenance
  • Native Vertex AI integration for enterprise workflows

Limitations

  • Per-image pricing (not token-based) — budget at image-count level
  • Strict safety filters can block legitimate creative asks
  • No native editing API as flexible as some rivals (Flux, GPT-Image)
  • Superseded for some creative uses by Imagen 4 where available

Use cases

  • Marketing imagery and concept art generation
  • Ad creative variations at scale on Vertex AI
  • Product visualisation for e-commerce
  • Editorial and presentation graphics

Benchmarks

BenchmarkScoreAs of
GenAI-Bench text-to-imageSOTA at launch2024-12
Human preference vs DALL-E 3preferred2024-12
Typography / text-in-imagestrong2024-12

Frequently asked questions

What is Imagen 3?

Imagen 3 is Google DeepMind's flagship text-to-image model, generally available from December 2024. It produces photorealistic images from text prompts, with improved typography and SynthID watermarking — accessible via Vertex AI, the Gemini API, and the consumer Gemini app.

How much does Imagen 3 cost?

Imagen 3 uses per-image pricing on Vertex AI and the Gemini API, typically a few cents per image depending on resolution and tier. Check Google's pricing page for the current per-image rate in your region.

What is SynthID and why is it in every Imagen 3 output?

SynthID is Google's invisible watermark for AI-generated media. It embeds a robust signal into pixels that survives cropping and compression, letting downstream verifiers identify Imagen-generated images. Every Imagen 3 image is watermarked by default.

Can Imagen 3 edit existing images?

Imagen 3 offers capability-limited image editing (masked inpainting, outpainting) through Vertex AI's Imagen tooling. For free-form editing, many teams combine Imagen 3 generation with separate editing pipelines.

Sources

  1. Google DeepMind — Imagen 3 — accessed 2026-04-20
  2. Google Cloud — Imagen on Vertex AI — accessed 2026-04-20