Google recently made available its latest text-to-image model—Imagen 2—on the Vertex AI platform. Initially announced during the I/O conference in May 2023, Google debuted the model on December 14, 2023. Imagen 2 has been built by several of Google’s subsidiaries including major contributions from Google DeepMind, which happens to be the firm’s core artificial intelligence laboratory. Promising better image quality along with the ability to create logos and add text to generated images, Google’s Imagen 2 signals the next step in the firm’s push to create better AI image generators. The training protocol—while not disclosed—has been focused on enhancing realism and strengthening AI’s generative capabilities in specific areas such as generating hands, words, and finer details.
Interestingly, Google AI has adopted a different approach by not launching Imagen 2 as an independent platform and instead making it available to approved Vertex AI customers via API. Design and digital creativity platforms such as Canva and Shutterstock have already jumped on the hype by adopting the new generative AI protocol into their existing workflows and offerings to their customers. Google has utilized advanced deep learning techniques integrated into highly specialized neural networks that have been primed to enhance the quality of image generation. The key details of the latest Imagen model are detailed in the forthcoming sections.
Essential Features of Google’s Imagen 2
Embedding text along with enhanced resolution are key improvements in Imagen 2.
Imagen 2 builds on the existing strides made by its predecessor and allows for high-quality image production. The text-to-image model is multilingual, with the ability to understand prompts in English, Chinese, Hindi, Korean, Japanese, Spanish, and Portuguese. Apart from precision, Google AI has also worked on enhancing the resolutions of output images and their aesthetic quotient. Clearly, Google is aiming to compete better with existing rivals in the AI image generation market, such as OpenAI’s Dall-E 3 and the independently created Midjourney platform. In addition to the text prompt, users can also provide context images to further fine-tune their outputs from the Imagen 2 model. The underlying natural language processing protocols are also adept at understanding and decoding long-form prompts and deriving key details from them.
Since firms and potential rivals like Amazon have also launched their own versions of image generators, such as Titan, Google intends to step up its game in the niche by making its offerings more relevant to clients such as advertisers and graphic designers. Since text can be overlaid in addition to the in-painting and out-painting of elements from existing images, it becomes far simpler to create interesting advertorial content that might otherwise take several work hours to perfect. While speeding up workflows, Google’s Imagen 2 can also overlay logos on images, essentially streamlining the process of branding. As for security, Imagen 2 uses Google DeepMind’s SynthID technology, which watermarks generated images down to the pixel and makes them resistant to attempts at editing out the AI safety features.
Why is Imagen AI’s Successor a Significant Development?
Google AI has not revealed the details of Imagen 2’s dataset.
Imagen 2 addresses significant flaws in image generation protocols and signals a change in how AI firms are approaching the problem. Since AI-generated content has become widely popular, key elements such as hallucinations and AI bias can only be addressed by extensive and pointed training—something Google has reportedly done with Imagen 2. The reduction of visual artifacts and enhancing facial expressions as well as features in AI-generated images has been a key concern in just about all AI image generators; however, Imagen 2 is claimed to have overcome this roadblock. Given that users spend a lot of time engineering prompts to get the desired output, the user’s ability to have more control over setting the model to better match their expectations is also important. The image quality scoring feature is an additional advantage, which allows users to rank the model’s output and further fine-tune their interactions with it.
That being said, there remain a number of outstanding concerns with Imagen 2 that pertain to the domains of AI privacy and copyright. Since Google has not publicly announced the nature of the data used to train the model, it is hard to determine the exact sources Imagen 2 uses for its image generation purposes. While it is a given that adversarial training protocols would be at the heart of Imagen 2’s technology, the lack of information on the dataset leaves both enthusiasts and experts wondering. The earlier model of Imagen relied on the publicly available LAION dataset; however, no such information has been made available for its successor. Since the issue of AI firms using potentially private information alongside copyrighted content has become a burning issue in the AI space, this might pose challenges to Google, which has so far not addressed the issue. The firm, however, does offer its customers indemnification benefits that protect them from any IP infringement claims leveled either on Google Cloud or Imagen 2’s services.
The Growing Footprint of AI-Generated Content
Google’s Imagen 2 might become one of the world’s most advanced AI image generation protocols.
With Google’s concerted push to improve its AI image-generation models, it is evident that AI content is bound to grow more popular as time progresses. With numerous firms, including Google, also taking into account responsible AI practices and including watermarks, it is important to draw distinctions between generative AI media and those produced by individual content creators. Google has also recently announced the arrival and availability of AI tools on YouTube, the world’s most popular video hosting platform. As AI becomes ubiquitous in the content space, attention will have to be paid to protecting the rights of artists and independent creators to ensure AI does not unfairly capitalize on original work to deliver monetary benefits to its source firms.
FAQs
1. Is Imagen 2 available?
Imagen 2 is generally available on Vertex AI for its customers. The protocol brings advanced image generation capabilities and better customizability to the text-to-image protocols.
2. Is Imagen 2 free?
While Imagen 2 allows free trials, it is not free and comes bundled with a subscription to Google’s Vertex AI platform.
3. How many languages does Google’s Imagen 2 support?
Imagen 2 supports over seven languages, including English, Chinese, Hindi, Japanese, Korean, Spanish, and Portuguese. More languages are set to be added to the platform in the early months of 2024.