Grok Imagine Video 1.5 (preview) is a video generation model developed by xAI, designed primarily for image-to-video (I2V) and video-to-video tasks. Released in late May 2026, the model is built on the Aurora engine, an autoregressive architecture that processes video clips sequentially to ensure high temporal consistency. Unlike many diffusion-based video models that generate frames simultaneously, the Aurora engine generates frames such that each informs the next, maintaining stability in subject identity, lighting, and camera trajectory throughout the clip.
Technical Architecture
The model is characterized by its unified transformer pass, which jointly processes video tokens and audio waveforms. This allows for native synchronization of visual motion with audio elements such as Foley sound effects, ambient acoustics, and lip-synced dialogue. The training of this model reportedly utilized a significant compute cluster involving approximately 110,000 NVIDIA GB200 GPUs, allowing for the processing of complex environments and photorealistic micro-expressions with high fidelity.
Capabilities and Specifications
grok-imagine-video-1.5-preview supports multiple aspect ratios, including 16:9, 9:16, and 1:1, at resolutions of 480p for drafting and 720p for final output. It generates clips at a standard frame rate of 24 fps with durations ranging from 5 to 15 seconds. Key features include video extension, which allows users to continue a sequence from the last frame of an existing clip, and reference-guided generation, which uses input images to maintain character or style consistency across multiple generations.
Prompting and Best Practices
For optimal results, the model requires a high-quality source image to establish scene geometry and lighting. Official guidance suggests front-loading key actions in the motion description; actions described early in a prompt are more likely to appear at the start of the sequence. Users are encouraged to specify camera behavior (such as dolly pushes or handheld moves) and ambient lighting details to guide the Aurora engine's sequential rendering process effectively. While the model excels at environmental and character consistency, it may exhibit subtle shifts in complex brand elements or dense typography during rapid camera movements.