Google unveils Imagen text-to-image diffusion model

Google Research unveiled Imagen, a text-to-image diffusion model, seven weeks after OpenAI released DALL·E 2. Google reported that Imagen achieved a lower (better) Fréchet Inception Distance score on the COCO benchmark than DALL·E 2 — 7.27 against 10.39 — and that human raters preferred its outputs over other contemporary models on both image fidelity and how closely an image matched its text prompt, rating Imagen samples as comparable to the COCO dataset’s own real photographs on alignment with caption text.

Architecturally, Imagen paired a large frozen language model — a T5-XXL text encoder that was not fine-tuned during image training — with a diffusion model to render the image. Google’s central finding was that scaling the text encoder mattered more to output quality than scaling the diffusion component itself, a result at odds with the more common assumption that image quality gains would come primarily from a bigger image model. Alongside Imagen, Google introduced DrawBench, a benchmark suite designed to probe capabilities existing benchmarks missed, such as rendering correct spatial relationships, counting, and unusual word combinations; Imagen led on DrawBench as well as on COCO FID.

Unlike DALL·E 2, which OpenAI opened to a waitlisted user base within weeks, Google did not release Imagen’s code, weights or a public demo, citing the risk of the model reinforcing social biases present in its training data and of enabling misuse such as disinformation. The gap between Imagen’s benchmark claims and its complete unavailability became a recurring point of criticism, since neither independent researchers nor the public could verify Google’s reported comparisons directly.

Imagen remained a research demonstration rather than a shipped product through 2022, while DALL·E 2, Midjourney and the openly released Stable Diffusion all reached users that year — leaving Google, despite claiming state-of-the-art results, on the sidelines of the text-to-image product race it had helped establish.