Luma adds Camera Motion Concepts to Ray2
Luma has released Camera Motion Concepts for its Ray2 video model, which it says lets creators reliably control and combine camera moves, each learned from one or a few examples, without degrading the model's quality.
Luma announced Camera Motion Concepts, the first release from a research track it calls "Concepts." The company describes this as a method for teaching generative models new ideas and controls from one or very few examples. The first set targets camera motion in its Ray2 video model and is available in Dream Machine starting today.
Luma says each Concept learns a specific camera move and reproduces it reliably across a wide range of shots and styles. It contrasts the approach with LoRA and fine-tuning in two ways.
The first is composition. Luma says multiple Concepts can be combined through natural language at inference time. That allows camera moves that may not be physically possible, and so could not be learned from training examples any other way.
The second is quality. Luma claims that adding Concepts does not compromise Ray2's quality, stylistic range or capabilities. It argues that existing customization methods "collapse model capabilities to the mode of the data" used to train them.
Camera Concepts also work with Ray2's other features: image-to-video, start and end keyframes, and loop.
The announcement does not include technical details, benchmarks or a list of the camera moves included. It also does not explain how Concepts are trained or represented. Luma says it will share more about the approach "in the near future" and release more Concepts-based capabilities for Ray2 and Dream Machine in the coming days and weeks. It is also asking users to suggest new Concepts.