Uni-1 by Luma

Uni-1 is Luma's inaugural unified understanding and generation model, designed to integrate reasoning and visual imagination within a single AI architecture. It represents a significant step towards multimodal general intelligence by jointly modeling time, space, and logic. The model operates as a decoder-only autoregressive transformer, processing text and images in a single interleaved sequence for both input and output. Uni-1 excels in performing structured internal reasoning before and during image synthesis, allowing it to decompose instructions, resolve constraints, and plan compositions effectively. This capability is demonstrated by its state-of-the-art performance on the RISEBench, a benchmark for Reasoning-Informed Visual Editing. Furthermore, the model's ability to generate images enhances its fine-grained visual understanding, leading to strong grounding and dense visual capabilities. The model supports a range of advanced visual tasks, including intelligent common-sense scene completion, spatial reasoning, plausibility-driven transformations, and maintaining temporal and causal consistency across evolving scenes. It also offers directable reference-guided generation, allowing users to control outputs using sketches, visual instructions, style, and structure, and supports multi-turn refinement.