Luma AI unveils Uni-1 autoregressive model that generates text and pixels simultaneously

Luma Labs launched its image generation model Uni-1 on March 23, marking the company's first publicly available model built on the Unified Intelligence architecture. Free trial access is now open on the official website, with API pricing announced and enterprise access channels rolling out gradually.

Architecture Shift: From Diffusion Models to Autoregressive

Uni-1 moves away from the prevailing diffusion model approach, opting instead for a decoder-only autoregressive Transformer. It arranges text and image tokens in an alternating sequence within a single sequence, completing inference and pixel generation in one forward pass.

Luma CEO Amit Jain explained that traditional solutions typically first use a language model for planning, then hand off to a diffusion model for generation, causing information loss between the two stages. Uni-1’s design aims to eliminate this gap.

Jain previously worked at Apple and contributed to Vision Pro engineering.

Capabilities: Reference Image Control and Cross-Style Generation

Uni-1 supports image generation guided by one or more reference images, preserving the subject's identity, posture, and composition. Official tests show stable performance in handling character consistency and portrait control in multi-reference image mode.

The model claims support for 76 visual styles, covering categories such as realistic photography, comics, and ukiyo-e.

In a demonstration, inputting "Draw an infographic of the Golden Gate Bridge" led the model to automatically plan the layout, generate a bridge structure diagram, and annotate data such as "1711 Meters," with the internal reasoning process visible in real time.

Benchmarks: Leading in Spatial Reasoning and Reference Generation

Data published by Luma shows that Uni-1 scored 0.51 on the RISEBench reasoning benchmark, higher than Google Nano Banana 2's 0.50 and OpenAI GPT Image 1.5's 0.46. Its spatial reasoning score reached 0.58, and logical reasoning 0.32, roughly double that of GPT Image.

On the ODinW-13 object detection benchmark, Uni-1 achieved 46.2 mAP, close to Google Gemini 3 Pro's 46.3.

In human preference Elo rankings, Uni-1 placed first in overall preference, style and editing, and reference generation, and second in text-to-image generation.

Pricing

API charges are token-based: $0.50 per million tokens for input text, $1.20 per million tokens for input images, $3.00 per million tokens for output text and thought chain, and $45.45 per million tokens for output images.

Converted to a per-image basis: text-to-image (2048px) costs approximately $0.0909, editing with a single reference image around $0.0933, and eight reference images about $0.1101.

VentureBeat reported that in enterprise scenarios with 2K resolution, Uni-1 costs 10% to 30% less than Google Nano Banana 2.

Background

Luma Labs previously focused on video generation products like Dream Machine (Ray3 series). On March 5, the company released the Luma Agents creative agent platform, also based on the Unified Intelligence architecture. Uni-1 is the first application of this architecture in a static image product.

Within hours of the release, related posts on the X platform garnered over 2.3 million views. Luma stated that video and audio versions will follow, though specific timelines have not been disclosed.

Try it at: lumalabs.ai/uni-1