According to 1M AI News monitoring, AI video company Luma Labs has released the image generation model Uni-1, which adopts a self-regressive Transformer architecture to simultaneously complete inference and pixel generation in a single model, taking a completely different technical route from mainstream diffusion models such as Midjourney and Stable Diffusion. The model does not rely on a two-stage process of "first understanding and then handing over to another model to draw," but rather continues to infer during the generation process, decomposing commands, parsing constraints, planning composition, and then rendering.
In the RISEBench benchmark specifically evaluating inference capabilities, Uni-1 has an overall score of 0.51, surpassing Google Nano Banana 2 (0.50) and OpenAI GPT Image 1.5 (0.46). The spatial reasoning score is 0.58, leading Nano Banana 2 at 0.47; the logical reasoning score is 0.32, more than twice that of GPT Image 1.5 (0.15). On the ODinW-13 object detection benchmark, Uni-1 scores 46.2 mAP, nearly matching Google Gemini 3 Pro (46.3). Luma stated that removing the generation training from the same model would result in a 2.3-point drop in understanding ability, demonstrating that learning to draw has in turn enhanced the model's visual understanding. In human preference Elo ratings, Uni-1 ranks first in overall quality, style editing, and reference image generation, only ranking second in text-to-image generation.
In terms of pricing, a 2K resolution text-to-image generation is priced at approximately $0.09 per image, lower than Nano Banana 2's $0.101 and Nano Banana Pro's $0.134, with a 10% to 30% discount for high resolution. The model is now available for free trial on lumalabs.ai, and API access will be gradually opened through a waiting list.
Luma Labs is headquartered in San Francisco with a team of about 150 people, previously known for its video generation tool Dream Machine. CEO Amit Jain previously revealed that the Luma Agents creative platform powered by Uni-1 has partnered with Publicis Groupe, Adidas, Mazda, and other brands, once compressing a $15 million, year-long ad project into 40 hours at a cost of less than $20,000.
