Meta licenses Midjourney's image and video technology

Background

Text-to-image and text-to-video models take a written prompt and produce a picture or a short clip. They are trained on very large collections of images or video paired with descriptions, and learn the patterns that link words to visual content; at use time the model starts from noise and gradually shapes it toward something that matches the prompt. Building such a model takes data, computing power and engineering time, which is why an organisation that wants the capability usually either builds it in-house, licenses it from a company that already has it, or buys that company.

Before the deal, Meta had gone some way down the first road. It had rolled out its own AI image tool, Imagine, into Facebook, Instagram and Messenger, and it had a video tool called Movie Gen^[2]. Image and video generation had become a familiar feature inside large social platforms — something users expect to find in an app they already open, rather than at a separate site.

The announcement

On 22 August 2025, Meta announced that it would license Midjourney's AI image and video generation technology^[1]. The deal was made public by Meta's Chief AI Officer, Alexandr Wang, in a post on Threads, in which he said that Meta research teams would collaborate with Midjourney^[1]. Financial terms were not disclosed^[1].

Two parts of the announcement are worth separating. One is the licence itself: permission for Meta to use Midjourney's image and video models. The other is the research collaboration Wang described, which points at joint work between Meta's researchers and Midjourney's team, not simply the handover of a finished product^[1].

Independence, and an earlier acquisition conversation

On the Midjourney side, David Holz said that the company remained independent and had no outside investors^[1]. TechCrunch noted that Midjourney was one of the few leading AI model developers that had never taken on outside funding^[2]. According to Upstarts Media, Meta had at one point talked with Midjourney about acquiring the startup^[2] — a conversation that ended, on the evidence of the August announcement, in a licence rather than a purchase.

Why a licence is not a purchase

A licence is permission to use someone else's technology, usually for a fee and on agreed terms, without transferring ownership of the company that built it. An acquisition works the other way: the company, its team and its technology all move inside the buyer.

The two arrangements mean different things for each side. For the licensor, licensing keeps the company, its team and its way of working intact, and turns the technology into an asset that can be paid for over time. For the licensee, licensing is faster than building from scratch and narrower than buying a company: it brings in one specific capability without taking on a whole organisation. A research collaboration sits between the two — it is more than buying access to a model, because it puts researchers from both sides to work together^[1].

Where the two companies stand

Read side by side, the facts of the announcement mark out each company's position. Meta had already shipped an image tool of its own in Imagine and a video tool in Movie Gen^[2], and it still chose to pay for access to someone else's image and video models^[1]. Midjourney was one of the few leading model developers that had never raised outside money^[2], and it chose to license rather than sell, remaining an independent company with no outside investors^[1].

The announcement set out a licence plus a collaboration: Meta gains the use of Midjourney's image and video generation technology, and the two companies' research teams are to work together, with Meta's Chief AI Officer announcing the arrangement^[1].