Midjourney must hand over video training data to the studios

A court-approved joint stipulation in the studios' copyright case against Midjourney requires the company to transfer the video datasets it used to train its models to the plaintiffs, handing Disney, Universal and Warner Bros. a file-level copy of the data behind a shipped AI video model.

The stipulation was signed by counsel for both sides on September 22, 2026, and U.S. Magistrate Judge A. Joel Richlin approved it as an order on September 29, 2026, according to the document filed in the U.S. District Court for the Central District of California. Aggregators that summarized the order, AIPopDB and The Inference, published their accounts on October 7 and 9, 2026. The filing is a joint stipulation, an agreement between the parties that the court signed off on, rather than a ruling imposed on Midjourney.

The protocol implements Midjourney's election under Rule 33(d) of the Federal Rules of Civil Procedure, made November 21, 2025 in response to Interrogatory No. 3, and responds to the plaintiffs' Request for Production No. 4, according to the order. It defines the Midjourney AI Models as the image models designated v4 through v8 and the company's video-generation model, and defines training data as the image and video datasets used to train them, including captions, metadata and retained embedding vectors.

The mechanics are spelled out to the day. Within seven days of the order, the plaintiffs designate Google Cloud storage buckets; Midjourney then transfers a complete copy of the datasets within seven days, or fifteen if the buckets sit in another region. The two sides split the transfer costs and the at-rest storage costs evenly. Midjourney must also produce a per-object manifest with checksums, a mapping of each directory to the model version trained on it, and a declaration signed under penalty of perjury. The data is designated Highly Confidential, and the parties bear their own analysis costs.

The plaintiffs are Disney, Universal and Warner Bros., whose suits are consolidated before Judge John Kronstadt. Disney and Universal filed in June 2025, and Warner Bros. followed that September, citing characters such as Batman and Superman. The studios say the case concerns Midjourney copying and distributing their films and shows and making derivative works from them. Midjourney argues that the studios use comparable AI tools inside their own businesses, which it says bears on its fair use and unclean hands defenses. "If Plaintiffs are doing the very thing they seek to punish, that evidence goes to the heart of Midjourney's fair use and unclean hands defenses," Midjourney attorney Bobby Ghajar said in court filings reported by VP Land.

Discovery has produced friction already. Law360 reported on June 16, 2026 that a California federal magistrate judge ordered the studios to produce some data on their own use of artificial intelligence but held that Midjourney's broader requests were irrelevant or shielded by work product privilege; that article is paywalled beyond its summary line. VP Land reported that the limit covered the studios' consumer-facing AI products. Midjourney asked Judge Kronstadt in early July 2026 to overturn that limit, and the reporting does not say how the request was resolved.

Midjourney is also producing sampled outputs in the same case: 385 jobs per character, with experts allowed to extrapolate counts to the full population, The Inference reports. Midjourney launched its video feature in June 2025, so the datasets at issue come from a product barely a year old.

The price of training on other people's work is partly on the record. The Inference, citing Music Business Worldwide, reports that Warner Music settled and licensed Suno in November 2025, that UMG licensed Udio a month earlier, and that all three majors took equity in Stability AI through a $76 million Series B in August, an investment rather than a damages figure. Suno's second suit claims 60,202 recordings at the $150,000-per-work statutory ceiling, which The Inference reports at just over $9 billion, against a company valued at $5.4 billion. None of those is a damages award. In the Midjourney case, Disney's and Universal's June 2025 complaint asks for damages and Midjourney's profits in an amount to be proved or, at their election, statutory damages of up to $150,000 for each infringed work, so no total is on the record.

For creators, the practical effect this week is small and the informational one is not. Nothing in the stipulation changes what Midjourney's tools generate or cost, and it grants no one else access to the files. What it does is put the datasets behind a shipped video model into the plaintiffs' hands with a manifest that maps each directory to the model it trained, a file-level record that outside estimates cannot produce. If the studios' experts turn that record into a claim, the bootstrapped Midjourney, which The Inference reports is profitable, citing The Information, would answer it from operating income rather than a reserve.

Under the order, the plaintiffs verify the copy by checksum against the manifest, Midjourney cures any mismatch within five business days, and unresolved defects go before Magistrate Judge Richlin.