Google has released TextFX, a new artificial intelligence image creation tool. TextFX is powered by Imagen2, the GenAI image model developed by the Google DeepMind team, and provides a prompt-based user interface to create and edit images.This is no different from tools such as OpenAI’s DALL-E3, Midjourney, Meta’s ImaginewithMetaAI and Microsoft Designer. However, the uniqueness of TextFX is "expression", which can be understood as a list of keyword suggestions, allowing users to try their creations and ideas in the "adjacent dimension".
Google wrote in a blog post: "Designed for experimentation and creativity, ImageFX lets you create images with simple text prompts and then easily modify them with new prompts using the Expression Chip."
Google claims it has taken steps to ensure TextFX is not used in unintended ways, such as by adding "technical safeguards" to limit "problematic output" such as violent, offensive and pornographic content. TextFX also sets up a prompt-level filter for "named people" (who may be public figures) - although Google doesn't make this particularly clear in its press materials.
"We've invested in the security of our training data from the beginning," Google said. "In line with our AI principles, we also conduct extensive adversarial testing and red teaming to identify and mitigate potentially harmful and problematic content."
As an additional security measure, Google marks images produced using ImageFX with a SynthID digital watermark, which is said to be highly resistant to image editing and cropping.
Google continued in the blog post: "SynthID watermarks are invisible to the naked eye but can be used for identification. With added insights in 'About this image,' when you see an image in Google Search or Chrome, you'll know whether it was likely generated by Google's artificial intelligence tools."
You can find ImageFX in AITestKitchen, Google's web app for artificial intelligence experimental projects.
Imagen2 extension
In related news today, Google said that starting this week, it will bring Imagen2 to more products and services, including the next-generation artificial intelligence search experience and the VertexAI series of artificial intelligence managed services.
Now, Imagen2 also supports the text-to-image feature in Google Ads and DuetAI in Workspace, the Google GenAI productivity product suite, and it has made its way into Google's SGE (Search Generation Experience). SGE began to provide users with image generation tools in Google Image Search in October last year, and now uses Imagen2 to generate images. Users can enter a prompt describing what kind of image they want, and SGE will return four results directly within the SGE conversational experience.
In VertexAI, Imagen2 is available to Google Cloud customers via API. Elsewhere, Imagen2 is now available via Bard, Google's AI chatbot.
Google explains: "With Imagen2, Bard can understand simple or complex prompts so you can generate a range of high-quality images. Simply enter a description -- such as 'Create an image of a dog riding a surfboard' -- and Bard will generate a custom, wide-range of visual images to help bring your ideas to life."
Google still hasn’t disclosed the data used to train Imagen2, which is not surprising. Whether GenAI vendors like Google can train models on public data (even copyrighted data) and then commercialize that model is an unresolved legal question.
Litigation is pending in court, and the vendors believe they are protected by the fair use doctrine. But it will take time for the dust to settle.
Google, meanwhile, is keeping silent on the matter as a precaution.