Luma has introduced Ray3.2, a significantly improved version of its Ray3 AI video generation model. The new version was developed in collaboration with creators from the entertainment, advertising, and gaming industries and is clearly aimed at professionals. The main change is much greater control over the output. Users no longer simply enter a text prompt and hope for the best. They can direct the scene frame by frame. According to the Luma platform, direction goes in and cinematic quality comes out.
Up to 16 frames in a single shot
The biggest advance over the previous version is the so-called multi-keyframe feature, which provides control at the individual frame level. Creators can place up to 16 keyframes in a single clip and use them to determine how the motion develops, what should change, and what should remain the same. Directors and production teams thus gain control over pacing, dramatic twists, and camera movement. The result can precisely match the storyboard and the client's expectations, which will be particularly appreciated by agencies working from detailed briefs.
The maximum length of generated shots has also increased. Ray3.2 can produce clips up to 20 seconds long at 1080p resolution. This makes it possible to build continuous scenes rather than just short fragments, a limitation that still hampers many competing tools.
Acting performances survive generation
The second major area of improvement is performance preservation. Enhanced performance tracking now covers posture and gestures, while facial expressions can be transferred across up to eight faces simultaneously, frame by frame. A studio can therefore take existing footage of an actor and transform the environment, costumes, or entire visual style while preserving the subtle nuances of the original performance.
This is also related to the redesigned Modify Video feature. It transforms existing footage: replacing a wall, world, or wardrobe while preserving the lighting and the acting performance. Here too, the limit is 20 seconds at 1080p.
Outputs ready for post-production
Ray3.2 supports formats that professional post-production teams actually use. The model generates HDR video with 16-bit color depth and can export 16-bit EXR frames in the ACES2065-1 color space. These are precisely the formats required by teams working on visual effects, compositing, and color grading. The generated shot therefore ceases to be merely an "AI experiment" and becomes a production-ready asset that fits into established workflows without losing dynamic range.
Higher quality comes at a price, however. HDR output costs twice as many credits as standard SDR, while the combination of HDR and EXR costs three times as many.
Adjusting a shot without reshooting
Another interesting new feature is the expanded Reframe tool. It allows users to modify a shot after it has been generated: changing the aspect ratio for different platforms, expanding the image, or replacing the background, all while preserving the scene's original lighting. Client feedback can thus be addressed with a targeted adjustment instead of the costly regeneration of the entire scene from scratch. For agencies preparing versions of a single campaign for different markets and formats, this represents significant savings in both time and money.
An API for developers for the first time
Ray3.2 is also the first model in the Ray series for which Luma has made complete control available through an application programming interface. Developers can therefore integrate video generation directly into studio software, internal production tools, or their own platforms. Luma designed the interface to fit into existing technology stacks, whether render farms or third-party applications.
This carries considerable weight in production environments. Teams there typically work with interconnected software tools rather than isolated applications. Media companies can therefore test the generative system directly within their own environments for asset management, editing, and approval. Platforms already working with Luma's technology include Adobe, Freepik, Envato, and Comfy.
Competition drives greater controllability
The emphasis on precise control reflects a broader shift in the AI video market. Vendors are seeking to attract film, advertising, and game productions, which require not only image quality but, above all, consistency and editability. Tools for these customers are increasingly judged by how well they fit into established production workflows and how effectively they handle rounds of revisions without requiring the scene to be rebuilt from scratch.
Competition in generative video is also intensifying. While consumer systems controlled solely through text prompts initially attracted attention, vendors are now under pressure to prove that their models can also meet the technical and operational demands of commercial production. Luma's response is precisely this combination of keyframes, performance preservation, and professional output formats. The company says that the future of AI in filmmaking does not lie in replacing creators' vision, but in enabling it. Storytellers have always known what they wanted to shoot. Now they are getting precise tools to do so.
Source: lumalabs.ai
