**[Huawei Hong Kong AI Lab Open-Sources Boogu: $400,000 to Train a 2K Image Model]** According to monitoring by Dongcha Beating, the Boogu team, led by Huawei Hong Kong Artificial Intelligence Lab and involving researchers from multiple universities, has open-sourced Boogu-Image-0.1. It is a 10-billion-parameter image generation and editing model, including Base, Turbo, Edit, and Edit-Turbo, supporting text generation in both Chinese and English. The foundational model was trained from scratch using only 208.62 million deduplicated images, with an estimated theoretical training cost of approximately $400,000. The paper claims that it ranks in the top tier of open-source models across multiple standard benchmarks, with some results approaching those of leading closed-source models. Base and Edit support up to 2K resolution, while Turbo can quickly generate 1K images in 3 to 4 steps. The system also interprets and rewrites user prompts before selecting a model based on task complexity, reducing unnecessary inference costs. The model weights, inference code, and some training methods are open-sourced under the Apache 2.0 license. [Original Link]