Huawei Hong Kong AI Laboratory open source Boogu: US$400,000 to develop a 2K graphics model

According to market news, the Boogu team, led by Huawei's Hong Kong Artificial Intelligence Laboratory and participated by researchers from many universities, has open-source Boogu-Image-0.1. It is a 10-billion-parameter image generation and editing model, including Base, Turbo, Edit and Edit-Turbo, and supports Chinese and English text generation. The basic model used only 208.62 million deduplicated pictures to train from zero, and the theoretical training cost was about US$400,000. The paper said that it reached the first echelon of open source in multiple standard evaluations, and some results were close to the leading closed-source model. Base and Edit support up to 2K, and Turbo quickly generates 1K pictures in 3 to 4 steps. The system will also understand and rewrite the user's prompt words first, and then select the model according to the difficulty of the task to reduce unnecessary reasoning costs. Model weights, inference codes and some training methods are open using the Apache 2.0 protocol.