Imagen

Free

Imagen by Google is a text-to-image diffusion model with an unprecedented degree of photorealism and a deep level of language understanding.

FreeFree tier

Inputs: textOutputs: image

Type

Open Source

Company

Google

Visit Website GitHub (40 stars)

About Imagen

Imagen is a text-to-image diffusion model developed by Google Research's Brain Team. It achieves unprecedented photorealism and deep language understanding by leveraging a large frozen T5-XXL language encoder to encode text, followed by a cascaded diffusion model that generates a 64×64 image and then upsamples it to 256×256 and finally 1024×1024 using text-conditional super-resolution models. Key innovations include the finding that scaling the pretrained text encoder improves performance more than scaling the diffusion model, a new thresholding diffusion sampler for large classifier-free guidance weights, and an Efficient U-Net architecture that is compute- and memory-efficient. Imagen achieves a state-of-the-art FID score of 7.27 on the COCO dataset without training on COCO, and human raters prefer its outputs over those of VQ-GAN+CLIP, Latent Diffusion Models, and DALL-E 2. To evaluate text-to-image models more thoroughly, the team introduced DrawBench, a comprehensive benchmark. Imagen is capable of generating highly detailed and imaginative scenes from complex text prompts.

Key Features

Uses a large frozen T5-XXL language encoder for text understanding

Cascaded diffusion model generating 64×64 → 256×256 → 1024×1024 images

Scaling text encoder size is more important than scaling diffusion model size

Thresholding diffusion sampler enables large classifier-free guidance weights

Efficient U-Net architecture for better compute and memory efficiency

State-of-the-art COCO FID score of 7.27 without training on COCO

Introduces DrawBench benchmark for comprehensive text-to-image evaluation

Human raters prefer Imagen over VQ-GAN+CLIP, Latent Diffusion Models, and DALL-E 2

Pros & Cons

Pros

Unprecedented photorealism and image quality
Deep understanding of complex text and prompt alignment
State-of-the-art performance on COCO benchmark
Efficient architecture compared to scaling diffusion models alone
Free research paper and benchmark available to the community

Cons

Primarily a research model; no public API or commercial product available
Requires significant computational resources for training and inference
May exhibit biases present in training data, as with other generative AI models

Best For

Generating photorealistic images from detailed text promptsCreative and imaginative scene generation (e.g., 'a brain riding a rocketship heading towards the moon')Research in text-to-image synthesis and diffusion modelsBenchmarking and evaluating text-to-image models using DrawBench

FAQ

What is Imagen?

Imagen is a text-to-image diffusion model developed by Google Research that creates photorealistic images from text descriptions.

How does Imagen work?

Imagen uses a large frozen T5-XXL language model to encode text, then a cascaded diffusion model generates a 64×64 image which is upsampled to 256×256 and finally 1024×1024 via super-resolution diffusion models.

What is the DrawBench benchmark?

DrawBench is a comprehensive benchmark introduced by Google to evaluate text-to-image models across diverse and challenging prompts.

What COCO FID score does Imagen achieve?

Imagen achieves a state-of-the-art FID score of 7.27 on the COCO dataset, without ever training on COCO.

Is Imagen available for public use?

Imagen is a research model with a published paper and benchmark, but no public API or product has been released.