Boogu: A 10B-Parameter Image Model Open-Sourced with Notable Training Efficiency

The research team led by Huawei's Hong Kong Artificial Intelligence Laboratory, in collaboration with several universities, has officially open-sourced Boogu-Image-0.1. This comprehensive suite for image generation and editing boasts 10 billion parameters and includes four variants: Base, Turbo, Edit, and Edit-Turbo, all supporting both Chinese and English text prompts.

Cost-Effective Training: $400K and 209M Images

A key highlight is its training efficiency. The team revealed that the Base model was trained from scratch on approximately 209 million carefully deduplicated images. The estimated computational cost for this training is around $400,000—a relatively modest figure in the era of multi-million-dollar AI model development.

Competitive Performance: On Par with Top Open-Source Models

Benchmark results indicate that Boogu's overall performance places it among the top tier of open-source image generation models. The research paper notes that its output quality in certain tasks and metrics approaches that of leading proprietary, closed-source models, demonstrating strong competitive potential.

Smart Features: Prompt Optimization & Adaptive Model Routing

The system incorporates practical intelligent features. It first parses and semantically refines user input prompts to enhance clarity and model comprehension. Furthermore, it automatically routes requests between different model variants (like Base and Turbo) based on perceived task complexity. This design aims to avoid using heavyweight models for simple jobs, thereby reducing unnecessary inference costs.

Specifications & Open-Source Details

  • Resolution: The Base and Edit models support generation up to 2K resolution. The Turbo variants prioritize speed, producing 1K images in just 3-4 inference steps.
  • Open-Source License: The project is released under the Apache 2.0 license. The release includes full model weights, inference code, and details on key training methodologies, facilitating community research and development.

The open-sourcing of Boogu introduces a new, cost-effective option in the AI image generation landscape. Its efficient training approach and intelligent workflow design could influence the direction of future open-source projects in this field.