Apple makes major AI advance with image generation technology rivaling DALL-E and Midjourney

Apple‘s Automatic learning search The team has developed a revolutionary AI system to generate high resolution images that could question the domination of diffusion models, technology feeding popular image generators as From And Media.

Advancement, detailed in a research article published last week, presents “Star screen“, A system developed by Apple researchers in collaboration with university partners who combine standardization flows with self -regressive processors to achieve what the team calls” competitive performance “with advanced diffusion models.

The breakthrough arrives at a critical moment for Apple, which faced criticize Above his difficulties with artificial intelligence. Monday Global Developer ConferenceThe company has only unveiled Modest AI updates to her Apple Intelligence Platform, highlighting the competitive pressure confronted with a company that many consider to be delayed in the AI ​​arms race.

“To Our Knowledge, this work is the first successful Demonstration of Normalizing Flows Operating Effectively at this scale and resolution,” Wrote the Research Team, which include Apple Machine Learning Researchers Jiatao Gu, Joshua M. Susskind, and Shuangfei Zhai, collaborators from institutions included UC Berkeley And Georgia Tech.

How Apple fights against Openai and Google in the wars of AI

THE Star screen Research represents Apple’s wider efforts to develop distinct IA capabilities that could differentiate its products from competitors. While companies love Google And OPENAI Dominated the headlines with their AI -generating advances, Apple has worked on alternative approaches that could offer unique advantages.

The research team has taken up a fundamental challenge in the AI ​​Image Generation: the scale of standardization flows to operate effectively with high -resolution images. Standardization flows, a type of generative model that learns to transform simple distributions into complex products, have traditionally been overshadowed by diffusion models and generative opponent networks in image synthesis applications.

“Starflow achieves competitive performance in the generation of conditional image and text generation tasks, approaching advanced diffusion models in the quality of the sample,” wrote the researchers, demonstrating the versatility of the system through different types of image synthesis challenges.

Inside the mathematical breakthrough which feeds the new Apple AI system

Apple’s research team introduced several key innovations to overcome the limits of existing standardization flow approaches. The system uses what researchers call a “deep mass design”, using “a deep transformer block [that] Capture most of the model representation capacity, supplemented by a few blocks of shallow transformers which are effective on the computer but substantially beneficial level. »»

The breakthrough also consists in operating in “the latent space of pre-trained self-employeders, which is more effective than direct modeling at the level of pixels”, according to the article. This approach allows the model to operate with compressed representations of images rather than gross pixel data, considerably improving efficiency.

Unlike diffusion models, which are based on iterative ranging processes, Star screen Maintains the mathematical properties of flow normalization, allowing “exact formation of maximum likelihood in continuous spaces without discretization”.

What Starflow means for future Apple iPhone and Mac products

Research comes as Apple faces increasing pressure to demonstrate significant progress in artificial intelligence. A recent Bloomberg analysis underlined how Apple Intelligence and Siri had a hard time competing with competitors, while the modest advertisements of Apple in WWDC underlined the challenges of the company in the space of IA.

For Apple, the exact likelihood training of Starflow could offer advantages in applications requiring precise control over the generated content or in scenarios where understanding of the uncertainty of the model is essential for decision -making – potentially precious for business applications and AI capabilities in Disvise that Apple highlighted.

Research shows that alternative approaches to distribution models can obtain comparable results, potentially opening new avenues for innovation that could play Apple forces in material integration and processing on devices.

Why Apple bets on university partnerships to solve its AI problem

Research illustrates Apple’s strategy of collaboration with the main university establishments to advance its AI capabilities. Co -author Tianrong ChenA doctoral student at Georgia Tech who did an internship with the Apple automatic learning research team, provides expertise in stochastic optimal control and generative modeling.

Collaboration also includes Ruixiang Zhang of the Mathematics Service of the UC Berkeley and Laurent Dinh, a researcher to the apprentice Brain Google And Depth.

“Above all, our model remains a flow of end -to -end standardization,” said the researchers, distinguishing their approach from hybrid methods that sacrifice mathematical towing for better performance.

THE Complete search document is available on arxivProviding technical details to researchers and engineers who seek to rely on this work in the competitive field of generative AI. While Starflow represents an important technical achievement, the real test will be whether Apple can translate these research breakthroughs in the type of AI functionalities oriented towards consumers who have made competitors such as chatgpt household names. For a company that has revolutionized entire industries with products like the iPhone, the question is not whether Apple can innovate in AI – it is if they can do it quickly.