UNI 1.1 is a foundational image generation model developed by Luma AI, characterized by its unique decoder-only autoregressive transformer architecture. Unlike standard diffusion models that iteratively denoise random signals, UNI 1.1 treats text and image data as a single interleaved sequence. This design enables "native reasoning," allowing the model to decompose instructions, resolve logical constraints, and plan visual compositions before and during the pixel synthesis process.
The model is engineered for high-fidelity tasks requiring precise instruction following, such as complex spatial reasoning and multi-object scene construction. Version 1.1 introduces significant improvements in photorealism, material accuracy, and cultural awareness. It is particularly noted for its ability to handle intricate character details and maintain consistency across different generations by integrating reasoning directly into the generative pipeline.
Key Capabilities
- Reasoning-First Synthesis: The model performs structured internal reasoning to ensure that generated elements follow real-world physics, spatial logic, and common-sense relationships.
- Advanced Typography: By processing text and visual tokens in a shared space, UNI 1.1 excels at rendering accurate, legible text and complex typography within images.
- Reference-Guided Generation: Users can provide up to nine reference images with assigned roles—such as character, style, or lighting—allowing for granular control over visual consistency.
- Flexible Aspect Ratios: The model supports nine native aspect ratios, ranging from ultra-wide 3:1 panoramas to ultra-tall 1:3 portrait formats, with output resolutions up to 2048 pixels.
Luma AI recommends using descriptive, natural language prompts rather than keyword-based strings. Detailed descriptions of the subject, environment, lighting, and atmosphere help the model's reasoning engine better interpret the user's intent and produce higher-fidelity results.