Huawei Hong Kong AI Lab Open Sources Boogu: $400,000 Trained 2K Face Image Generation Model

According to the latest report from TrendForce, the Boogu team, led by Huawei's Hong Kong AI Lab and involving researchers from multiple universities, has open-sourced Boogu-Image-0.1. This is a 10 billion parameter image generation and editing model, including Base, Turbo, Edit, and Edit-Turbo, supporting both Chinese and English text generation.

The Base model was trained from scratch using only 2.0862 billion de-duplicated images, with a theoretical training cost of around $400,000. The paper claims that it has reached the top echelon of open-source models in multiple standard benchmarks, with some results approaching those of leading closed-source models.

The Base and Edit models support up to 2K resolution, while Turbo can quickly generate 1K images in 3 to 4 steps. The system first understands and rephrases the user's prompts, then selects the model based on task difficulty to reduce unnecessary inference costs.

The model weights, inference code, and some training methods are released under the Apache 2.0 license.