Luma AI opens Dream Machine, its text-and-image video generator

Luma AI has made Dream Machine available to the public today, a video generator that creates clips from written instructions or an image. The launch matters less because it introduces a new idea than because it brings a model comparable in ambition to offerings such as Sora and Veo to an open website—models that had remained available only to a limited audience.

The tool generates five-second sequences, equivalent to 120 frames at 24 frames per second. Luma says its model can produce those 120 frames in about 120 seconds, a turnaround designed to let users try several versions of the same scene without waiting hours for each result.

A video model to try, not just watch

Dream Machine accepts a description of a scene—for example, a shot of a person walking down a rainy street—and produces a short video. It can also start from an image, a particularly useful feature for anyone who wants to preserve the look of a character, product or illustration while adding motion.

Luma describes Dream Machine as a transformer, the family of models that made large language systems popular and is now also being applied to image, audio and video. In this case, it was trained directly on video to learn not only what an object looks like in a frame, but how it should change from one moment to the next.

That difference is central. Generating a static image is already relatively common; video requires preserving the identity of characters, lighting, perspective and the trajectory of objects across dozens of consecutive frames. When the model fails, hands change shape, objects warp and movements defy physics.

The company showcases scenes with moving cameras, people, animals and complex environments. They are striking demonstrations, but it is important to distinguish between a clip selected for its strong result and everyday use: consistency still depends heavily on the prompt, the complexity of the scene and how many attempts the user is willing to make.

Generative video moves beyond closed testing

The announcement comes four months after OpenAI unveiled Sora, its video model, which had raised expectations with long, visually convincing sequences while remaining restricted to a limited group of evaluators. Google also introduced Veo in May, as part of a private program for selected creators.

Luma’s decision changes the competitive landscape because it turns comparison into something practical. Anyone can write a prompt, upload an image and assess both the system’s visual quality and its flaws. That openness could attract creators, agencies and product teams that need rapid prototypes, but it also exposes the model to scenes that are far less controlled than those in a presentation.

This is not Luma’s first foray into visual creation. The company made a name for itself with mobile-based 3D capture tools and Genie, a model for creating 3D assets. Dream Machine extends that push into a much broader consumer format: short-form video designed for social media, advertising, presentations and proof-of-concept work.

Faster ideation, not a camera replacement

For video professionals, the immediate use is not to replace an entire shoot. Clips last only a few seconds, and it is still difficult to request highly precise actions, continuity between shots or a specific performance. Its value lies in the early stages: visualizing an idea, exploring a style, preparing an animatic—an audiovisual sketch of a sequence—or creating short assets that can later be edited with conventional footage.

Image generation also opens a clear commercial avenue. A retailer can animate a product photograph; a studio can turn concept art into an atmosphere test; a communications team can prepare several versions of an ad before investing in production. In all these cases, the main savings come from speeding up exploration, not from guaranteeing a final result ready for publication.

The expansion of these tools also underscores two problems the industry still needs to solve: the provenance of training data and the ease of creating realistic-looking images. As the models become available to more users, safeguards against impersonation and mechanisms for identifying generated content will become increasingly important.

Dream Machine puts Luma among the companies competing to define generative video. The real test starts now: whether its speed and public access outweigh the limitations that models still face when representing motion, causality and visual continuity.