Google’s AI artist brings words to life with surreal imagery

Google has developed its own AI-powered text-to-image creator that allows users to input a sentence to generate a whimsical, photorealistic drawing.

Photo credit: Imagen

Imagen uses a text encoder to “understand” inputted words and uses a diffusion model to convert this understanding into imagery. The AI program is trained on large, uncurated datasets scraped from the web.

To assess the quality of the photos that are generated, Google developed an in-house benchmark called DrawBench, which consists of a list of 200 prompts. The company found that human raters preferred Imagen’s output compared to competitors like the DALL-E 2, VQ-GAN+CLIP, and Latent Diffusion models.

Chart showing Imagen’s performance over its competitors / Photo credit: Imagen

The company has not released code for Imagen due to fear of algorithmic bias, as its generator relies on a largely unfiltered dataset. By inputting inappropriate or misguided text, users could potentially create images containing derogatory stereotypes, pornography, or racism.

“Imagen relies on text encoders trained on uncurated web-scale data, and thus inherits the social biases and limitations of large language models. As such, there is a risk that Imagen has encoded harmful stereotypes and representations, which guides our decision to not release Imagen for public use without further safeguards in place,” Imagen researchers wrote.

Photo credit: Imagen

Though Imagen is not publicly accessible yet, its website does offer a limited interactive demo, where users can press pre-selected phrases to generate their own photos.

AI-generated art has become popular in recent years, with researchers striving to build fully creative autonomous programs to rival human artists. Google has made a previous attempt in the field, launching its algorithm-vision generator DeepDream in 2015.

See also: Series SEA: Who’s investing in the region’s AI startups?