Google is releasing the second version of Imagen, its AI model that can generate and edit images given a text prompt, but at least to Google Cloud customers using Vertex AI who have been cleared for access.

But the company isn't saying which data it used to train the new model — nor introducing a way for creators who might've inadvertently contributed to the dataset to opt out or apply for compensation.

Dubbed Imagen 2, Google's advanced model — which was rolled out to the public in preview at the tech firm's I/O conference in May — was developed using technology from Google DeepMind, Google's flagship AI lab. Compared to the first-gen Imagen, it's "significantly" improved in terms of image quality, Google claims (the company bizarrely refused to share image samples prior to this morning), and introduces new capabilities, including the ability to render text and logos.

"If you want to create images with a text overlay — for example, advertising — you can do that," Google Cloud CEO Thomas Kurian said in a press briefing on Tuesday.

Text and logo generation brings Imagen in line with other leading image-generating models, like OpenAI’s DALL-E 3 and Amazon’s recently launched Titan Image Generator. In two possible points of differentiation, though, Imagen 2 can render text in multiple languages — specifically Chinese, Hindi, Japanese, Korean, Portuguese, English and Spanish, with more to come sometime in 2024 — and overlay logos in existing images.

"Imagen 2 can produce … lettermarks and abstract logos… and has the ability to layer these logos on products, clothing, business cards, and other surfaces," Vishy Tirumalasetty, lead of generative media products at Google, explained in the blog post provided to TechCrunch ahead of today's announcement.

Thanks to "novel training and modeling techniques," Imagen 2 can also understand more descriptive, long-form prompts and provide "detailed answers" to questions about elements in an image. These techniques also expand Imagen 2's multilingual understanding, Google says — allowing the model to translate a prompt in one language to an output (e.g. a logo) in another language.

Imagen 2 uses the technique developed by DeepMind called SynthID, to embed invisible watermarks on images produced by it. Of course, to detect these watermarking – which Google touts as resistant to image edits involving compression, filters and color adjustments – is dependent on another Google-provided tool that third parties don't have access to. But perhaps it will help address some concerns when politicians begin to voice their worries over the explosion of AI-generated disinformation on the web.

Google declined to disclose the training data on which it relied when training Imagen 2, which comes as something of a disappointment. However it's not a complete shock. How an open legal question remains as to whether GenAI vendors like Google can train a model on publicly available, even copyrighted, data and then turn around and commercialize that model.

Relevant lawsuits make their way through the courts, but vendors are arguing that they are covered under fair use doctrine. It will be quite a few years yet before the dust settles.

Meanwhile, Google's playing it safe by keeping mum on the issue-a reverse in strategy of the first-gen Imagen, where it revealed that a version of the public LAION dataset was used in training the model. Its well-known for problematic content — such as private medical images, copyrighted artwork, or photoshopped celebrity porn, and that is clearly not the best look for Google.

Some AI image generator companies, such as Stability AI and — as recently as a few months ago — OpenAI, permit creators to opt out of training datasets if they wish. Others, such as Adobe and Getty Images, are setting up compensation schemes for creators — though not often generous or transparent ones.

Google — and, to be fair, a number of its competitors, including Amazon — does not provide a way to opt out or offer creator compensation. That's not likely to change anytime soon, it seems.

Google does have an indemnification policy that provides protection to eligible Vertex AI customers against copyright claims both arising from Google's use of training data and Imagen 2 outputs.

For corporate customers and devs, regurgitation or when a generative model spits out a mirror copy of training example is rightly a concern. Researchers find that the first-gen Imagen wasn't immune to this phenomenon; spitting out identifiable photos of real people, copyrighted work by artists and more when prompted in specific ways.

Not surprisingly, a recent poll of Fortune 500 companies by Acrolinx revealed that nearly a third have intellectual property as their number one concern over the use of generative AI. Another poll revealed nine out of 10 developers "heavily consider" IP protection when deciding whether to use generative AI.

It's a concern Google hopes its newly expanded policy will address. (Google's indemnification terms didn't previously cover Imagen outputs.) As for the concerns of the creators, well… they're out of luck this go-round.