Apple has recently released two key studies in the field of machine learning, showcasing its latest advancements in using large language models (LLMs) to simplify UI development and enhance image security reviews.
SQUIRE: Say Goodbye to "Black Box" Design
To address the pain points of "poor controllability and difficult fine-tuning" in the AI-generated UI process, Apple introduced a system called
Differing from tools that directly generate static designs, SQUIRE first generates a customizable component tree. Developers can modify fonts, add layers, or replace specific elements like building blocks before delivering the final code. This "white-box" process transforms the time-consuming "prompt trial-and-error cycle" into an intuitive interactive collaboration. Once the prototype is confirmed, SQUIRE can convert it into HTML and CSS code with one click. In testing with 11 front-end developers, the tool received high praise for its ease of use and controllability.
SafetyPairs: Giving Visual AI a "Protective Suit"
In terms of image generation safety, Apple proposed the SafetyPairs framework. By using counterfactual image generation technology, researchers created 1510 pairs of extremely similar images that differ in key features (for example: one image shows a building normally, while another shows the building on fire).
This dataset aims to precisely identify weaknesses in visual language models' safety ratings. Through this "spot the difference" type of training, Apple is able to develop more efficient protective models. This technology is expected to further enhance the safety level of local AI tools such as
In the latest
