UNI-1 (full name Unified Intelligence) is a multimodal reasoning model whose core feature is that it possesses both visual understanding and image generation capabilities, integrating the two into the same decoder-only Transformer architecture. It is officially described as not treating understanding and creation as two separate problems, but rather handling them uniformly with a single model.
Main capabilities
- Reasoning-driven image generation: supports commonsense scene completion, spatial reasoning, and plausibility-based transformations, emphasizing not only following prompts but understanding them.
- Multi-reference image composition: can accept multiple photos and merge people, objects, or environments from different source images into a coherent new scene.
- Temporal sequence and aging simulation: through a one-sentence prompt, generates a character's evolution sequence from childhood to old age under a fixed camera angle, automatically handling causal logic such as bodily aging and family changes.
- Multi-turn conversational refinement: can maintain context across multiple turns of dialogue and gradually adjust the subject, supports converting images into more than 76 artistic styles, accepts sketches and visual instructions as input, and transfers identity, pose, and composition from reference photos.
- Culturally aware visual intelligence: oriented toward scenarios such as global aesthetics, memes, and comics, emphasizing understanding of subtle contexts that general models easily overlook.
- In-image text rendering: officially claimed to generate complex characters, including Chinese idioms and non-Latin scripts, with very few typesetting errors, surpassing most competitors in text rendering accuracy.
How to use According to the official website, the process for using UNI-1 Studio is:
- Describe the image: enter a prompt in the input box, select Generate to open the studio and load the description.
- Set up the image: add reference images as needed, choose the aspect ratio and output format, and log in.
- Generate and download: view the credit consumption, generate the image, then preview and download the result. The official website states that a free trial is available and you can start without an API key; API access is gradually opening up, and enterprise options require contacting support.
Architecture and technical claims UNI-1 adopts a purely autoregressive unified paradigm, using a pure decoder-only Transformer architecture consistent with GPT-type language models, representing text and images as a single interleaved sequence that serves as both input and output, thereby performing structured reasoning before and during image synthesis. It is officially claimed to differ from traditional diffusion models and the chained pipeline of a "language model + separate image generator," and to eliminate the context gap caused by handoffs between modalities.
Benchmarks and pricing The official website claims that UNI-1 surpasses Google's Imagen 3 and OpenAI's GP on reasoning-type benchmarksT Image 1 is close to Google's Gemini 3 Pro on object detection; it leads in all four dimensions of RISEBench (reasoning-driven visual editing): temporal, causal, spatial, and logical; in human preference Elo, it ranks first in overall quality, style and editing, and reference-based generation, and second in text-to-image. On pricing, the official statement says that at 2K resolution the API text-to-image cost is about $0.09 per image, roughly 10% to 30% lower than similar Google models. The official website currently uses account credits for billing: 10 credits per image under the 1K studio setting. The above API prices are reference prices for the model; for website plans, please check the pricing page.
Use cases Applications listed officially include: infographics after structured reasoning, charts, Morse code, and multilingual text generation; reference-guided editing, sketch-to-image, and text-to-image; multi-turn creative workflows; and end-to-end creative processes combining text, images, video, and audio. The official website says that a one-year, $15 million international advertising campaign was converted into low-cost multilingual localized versions within 40 hours and passed strict internal review; and says it is used by brands such as Publicis Groupe, Serviceplan, Adidas, and Mazda for agency-grade production.
Limitations and information not yet provided Some capability statements, benchmark rankings, prices, and customer cases all come from official self-descriptions, with no third-party independent verification information provided; the specific available regions, complete plans and pricing details, API launch schedule, concurrency and quota limits, etc. are not explained in the provided materials.
