Adobe's Firefly video model claims licensed-only training

The shot

On 12 February 2025, Adobe announced the Firefly Video Model, a text-to-video and image-to-video generator built into the company's Firefly web application and, in beta, into Premiere Pro's Generative Extend. The announcement names the model the industry's first commercially safe AI video generation model, aimed at enterprise and agency use rather than hobbyist generation; the release quotes marketing and creative leads at Deloitte Digital, Stagwell, and other agencies describing early use in production ideation.

What the documents show

Adobe's own announcement is specific about training data: Firefly is only trained on content that Adobe has permission to use, which includes licensed content from Adobe Stock and public domain content, never on Adobe customer content. The company's separate product page, a living document, repeats this claim and adds that Adobe does not train Firefly or partner models on Creative Cloud subscribers' personal files. The commercially safe language is Adobe's own characterization, tied to that training-data policy and to the model's output carrying Content Credentials for provenance, and it is a vendor claim about legal risk exposure, not an independently adjudicated finding; no court ruling or third-party audit is cited in either document. The announcement is a promotional launch statement, though the training-data description within it is a specific, checkable factual claim rather than a vague assurance.

The workflow

A user generates a clip inside the Firefly web application from a text prompt or a still image, then can move that output into Photoshop, Premiere Pro, Lightroom, or Adobe Express for further editing, according to Adobe's product page. Inside Premiere Pro specifically, the same underlying model powers Generative Extend, letting an editor pull a Firefly-generated clip into an existing sequence and treat it like any other timeline asset, including trimming, grading, and adding sound.

What the tool does not change

Neither document claims the training-data policy guarantees a specific output is free of resemblance to a third party's copyrighted work, a person's likeness, or a trademark; the claim is about what data trained the model, not a guarantee about every generated frame. Clearing a specific shot for a specific production, checking a client's own risk tolerance, and deciding whether Content Credentials metadata satisfies a broadcaster's or distributor's delivery requirements remain decisions for the production and its counsel, not settled by Adobe's own description of the model.

• Does commercially safe as Adobe defines it match what this production's insurer or distributor requires?
• Is the generated clip's Content Credentials metadata preserved through the full delivery pipeline?
• Would this specific use case benefit from independent legal review despite the vendor's own assurance?

Adobe's training-data claim is specific enough to be useful, and specific enough to be checked against future disclosures. Treating commercially safe as Adobe's own defined term, rather than a general legal verdict, keeps the claim in its proper place.