Boogu-Image-0.1 matches closed-source models with just $400K training cost

A team of 33 researchers from multiple institutions has introduced Boogu-Image-0.1, a family of open-source unified multimodal models designed for both understanding and generation. The family includes four variants: Base (standard text-to-image), Turbo (fast inference), Edit (instruction-based editing), and Edit-Turbo (fast editing). Despite being trained on a modest dataset of 208.62 million unique images with an estimated compute cost of just $400K, the models deliver competitive performance against leading closed-source systems such as Nano-Banana-Pro and GPT-Image-2. Key innovations include targeted improvements in model understanding, data quality, and training pipelines, combined with agentic inference-time scaling that boosts generation and editing quality under constrained budgets.

On standard benchmarks, Boogu-Image-0.1 consistently matches or outperforms other open-source models and narrows the gap with closed-source alternatives. The model also supports bilingual text rendering in both Chinese and English, a feature typically absent from many open-source offerings. The authors attribute their success to careful data curation and efficient scaling strategies rather than brute-force compute. They emphasize that their findings challenge the assumption that state-of-the-art multimodal generation requires massive proprietary datasets and budgets.

Boogu-Image-0.1 is fully open-source under the Apache 2.0 license, with weights, code, and training recipes publicly available. This release aims to advance the open ecosystem for unified multimodal understanding and generation, providing a strong baseline for researchers and developers who need high-quality image generation without reliance on expensive APIs or closed models. The project's code is hosted on GitHub, and the paper includes extensive ablation studies and practical discussions for the broader community.