Oracle AI Media Models Enter US Beta on OCI

Oracle is letting a small group of US companies try three AI models it built itself: one that generates video, one that generates images and one that edits them. The company unveiled Oracle AI Media Models at its AI World conference in Las Vegas on 27 October, in a post on its AI and data science blog by Sujith Ravi, its group vice president for AI foundations, who calls them “Oracle’s first family of foundation models”.

Until now, Oracle’s AI World news has been about business software and the firms that sell it. This launch is different: Oracle is making the models, and its main pitch is where the training data came from. Oracle says the models were “trained from scratch on rights-cleared data”. That is a claim about copyright as much as about technology, and Oracle has not published the details that would let anyone test it.

Three separate tools, reached through Oracle’s cloud

Each model does one job. Video Generation turns a text prompt into a clip. Image Generation does the same for still pictures. Image Editing takes “one existing image and one text instruction to add, remove, or modify elements”, such as adding a canoe to a lake or changing its colour, in Oracle’s own example. Oracle’s product page says prompts are in English.

Developers reach the models through managed endpoints on Oracle Cloud Infrastructure (OCI), and Oracle leaves the workflow to the customer. “The organization decides how those results move through creative review, asset management, and approval,” the blog post says. Oracle is also adding tools to search a company’s own image and video archive “semantically by scene, subject, or action”, along with services to prepare data and customise the models.

Access is tight. The models are in a beta programme for “select customers”, and the product page says access “is limited to selected US companies”. The models are “currently available in the U.S. only”, so UK studios and agencies cannot use them yet, and Oracle has given no date for other countries. There is also no pricing, no model size, no benchmark result and no date for general availability.

Shutterstock is exploring, not committed

The only outside name in the launch is Shutterstock, which Oracle calls “one of our data partners”. Oracle does not say whether Shutterstock’s images or footage were used to train the models. Daniel Mandell, Shutterstock’s senior vice president of data licensing and AI, is quoted in Oracle’s post. “For generative AI to become a trusted tool in the entertainment industry, the quality and integrity of the data used to train models are critical,” he said, adding that the company is “exploring” the generation API and search tools for its own collections.

The models also sit inside a bigger product. In a press release the same day, Oracle announced Oracle Media Fabric, which bundles the media models with a production suite called Oracle Creative Studio and a video streaming service called Oracle Video Edge. Oracle says Media Fabric “includes safeguards to help protect customers’ intellectual property (IP) and reduce the risk of generated content infringing the IP rights of others”, but the release does not say how those safeguards work. Media Fabric is “planned to be available within the next 12 months”.

“Creating content and delivering it to viewers were separate processes,” said Oracle chief executive Clay Magouyrk, who presented Media Fabric as a way to bring the two together.

The launch follows Oracle’s plan, announced a day earlier at the same event, to let ChatGPT and other AI tools run work in its Fusion business apps, and its $1 billion pledge to its partner network. It also comes days after Canva released Craft, its first in-house image generation model, and on the same day Apple said its new Creator Studio assistants are built with OpenAI as well as its own models.

Oracle is selling legal comfort before it has shown quality

The interesting part of this launch is not that Oracle can generate a video of a mountain lake. It is that Oracle leads with the data rather than the output. Mandell’s own words point to why: for media companies, the integrity of training data is the price of entry. Lawsuits over AI training, tracked in TSN’s AI copyright lawsuits hub, keep coming, most recently a claim that Anthropic’s Claude was trained on 274 Elvis songs. A studio’s lawyers may care more about where a model’s training data came from than about how sharp its pictures look.

But “rights-cleared” is only as strong as the evidence behind it. Oracle has not said what the data is, how much there is, who licensed it or on what terms. For anyone buying with a compliance eye, the useful questions are the dull ones: whether Oracle will set out the data’s origin in writing, whether its IP safeguards come with contract terms, and who carries the risk if a generated clip copies someone else’s work. Neither document answers them.

There is also a test of confidence built into Oracle’s own design. Creative Studio is “model-agnostic” and can call “leading third-party models”, so Oracle’s models will have to win work inside Oracle’s own product. With no benchmarks and a closed beta, nobody outside Oracle can yet say whether they will.

What to watch

The next signals are a price, a date for wider access and any detail on the training data. Shutterstock’s word is “exploring”; a signed deal would mean more. The broader Media Fabric platform is due within 12 months, by Oracle’s own schedule, and until then the models stay in a US-only beta.