Uni-1 by Luma

Uni-1 integrates both visual understanding and image generation capabilities within a single model, allowing for a cohesive approach to multimodal intelligence.

The model can perform structured internal reasoning processes, including decomposing instructions, resolving constraints, and planning composition, before and during image synthesis.

Uni-1 achieves state-of-the-art results on RISEBench, demonstrating its capacity for complex visual editing that requires temporal, causal, spatial, and logical reasoning.

Learning to generate images significantly improves Uni-1's fine-grained visual understanding performance, enabling reasoning over regions, objects, and layouts.

The model can complete scenes with common-sense understanding, applying spatial reasoning and plausibility-driven transformations.

Uni-1 maintains consistency across time and evolves scenes through coherent motion and event progression, as demonstrated by features like 'Temporal Layering' and 'Causal Game'.

Users can guide generation using various references, including sketches, visual instructions, style, and structure, allowing for precise control over outputs.

Uni-1.1 is available via API, offering a 'pay per image' model suitable for prototyping and production pipelines, with mentions of free tiers becoming viable.