Imagen

Overview

Imagen is a text-to-image diffusion model developed by Google Research's Brain Team. The system is designed to synthesize photorealistic images from natural-language text, combining an unprecedented degree of photorealism with deep language understanding. Imagen leverages large transformer language models, specifically a frozen T5-XXL encoder, to encode input text into embeddings. A conditional diffusion model maps these embeddings into a 64×64 image, and text-conditional super-resolution diffusion models progressively upsample the output to 256×256 and finally 1024×1024. This cascaded design enables high-resolution generation while maintaining a strong alignment between the text prompt and the resulting image.

One of Imagen's key discoveries is that generic large language models pretrained on text-only corpora are surprisingly effective at encoding text for image synthesis. Scaling the pretrained text encoder yields larger gains in sample fidelity and image-text alignment than scaling the image diffusion model. The architecture also introduces an Efficient U-Net that is more compute-efficient, memory-efficient, and converges faster. A new thresholding diffusion sampler allows the use of very large classifier-free guidance weights, further improving sample quality. These technical innovations allow Imagen to achieve a state-of-the-art COCO FID score of 7.27 without ever training on the COCO dataset, and human raters find Imagen samples to be on par with the COCO data itself in terms of image-text alignment.

To enable a comprehensive evaluation of text-to-image models, the team introduced DrawBench, a challenging benchmark with prompts designed to test compositionality, cardinality, spatial relations, long-form text, and rare words. In side-by-side human preference comparisons, Imagen is consistently preferred over recent methods including VQ-GAN+CLIP, Latent Diffusion Models, and DALL-E 2, both in sample quality and image-text alignment. The model also outperforms prior state-of-the-art approaches on COCO FID, as shown in the official table comparing AttnGAN, DM-GAN, XMC-GAN, Make-A-Scene, DALL-E, GLIDE, and DALL-E 2.

Imagen is part of a broader family of text-to-image research at Google Research, including Parti, Imagen Video, and Imagen Editor. It is intended for research and demonstration purposes, with potential applications in creative ideation, content generation, and benchmarking. However, the team has deliberately not released code or a public demo due to ethical challenges and societal risks. Imagen inherits social biases from uncurated web-scale training data, and its image fidelity degrades when depicting people. Future work will focus on responsible externalization frameworks and improving bias evaluation methods for text-to-image models.

Core capabilities

Cascaded Diffusion Architecture

Generates images via a base 64×64 diffusion model and two text-conditional super-resolution stages up to 1024×1024.

Frozen T5-XXL Text Encoder

Uses a large pretrained language model to deeply understand text prompts and encode them for image synthesis.

Efficient U-Net

A compute- and memory-efficient architecture that converges faster than previous diffusion backbones.

Thresholding Diffusion Sampler

Enables very large classifier-free guidance weights for improved sample quality and image-text alignment.

State-of-the-Art COCO FID

Achieves a 7.27 FID score on COCO without ever training on the COCO dataset.

DrawBench Benchmark

Introduces a comprehensive benchmark testing compositionality, cardinality, spatial relations, long-form text, and rare words.

Human-Preferred Generation

Outperforms DALL-E 2, Latent Diffusion Models, and VQ-GAN+CLIP in side-by-side human evaluations.

Research Family Integration

Part of the Imagen family including Imagen Video and Imagen Editor, enabling extended image and video generation research.

Use cases

Generating photorealistic images from complex descriptive prompts

Creative ideation and rapid prototyping of visual concepts

Producing high-resolution artwork from natural language descriptions

Evaluating text-to-image models with challenging DrawBench prompts

Researching scaling effects of language models in image synthesis

Extending to video generation and image editing with the Imagen family