Updates·February 2, 2024, 00:52

Google's Bard chatbot can now generate images from your text

AI-generated from the sources listed.

Google has given its Bard chatbot the ability to create images from a description. The feature is free, and the images get an invisible watermark so they can be recognized as AI-made.

Google has added image generation to its Bard chatbot. Users type a description of the image they want, and Bard creates images from the text. The feature launched on February 1, 2024.

According to Engadget, image generation is free and available in English in most countries. The sources do not say which countries are excluded.

New image model behind it

The images are created with Google's Imagen 2 image model, developed by Google DeepMind. Google says Imagen 2 produces its best images yet, with fewer visible flaws, especially in hands and faces. That is Google's own assessment, and the sources have not tested it independently.

Watermark is meant to expose AI images

All images from Bard get a digital watermark called SynthID, developed by Google DeepMind. It is embedded directly in the image's pixels and cannot be seen with the naked eye, but a computer can detect it.

Google also plans to use the "About this image" feature, which is meant to show that content was made with AI, in Google Search and Chrome.

According to Engadget, Google has also introduced safeguards: checks on training data, technical guardrails and what is known as adversarial testing, where the company probes whether the system can be tricked. There are also filters against images of named people, and the filters are meant to block violent, offensive and sexually explicit content as well as deepfakes. How well the filters work is Google's own claim and has not been independently tested.

Also a new experimental tool

At the same time, Google launched ImageFX, an experimental image tool in AI Test Kitchen. It is built on Imagen 2 and has so-called "expressive chips" that make it easier to tweak a description. At launch, it was available only in English in the US, Kenya, New Zealand and Australia. Images from it also get the SynthID watermark as well as IPTC metadata, which is information stored in the image file.

On the same day, Gemini Pro in Bard was expanded to all languages and regions where Bard is available. Engadget cites more than 40 languages and more than 230 regions.

According to Engadget, Imagen 2 is also used in Search Generative Experience, Google Ads, Duet AI in Workspace and Vertex AI.