Boogu-Image, a unified open-source model for generating and editing images

Boogu-Image-0.1 is a new open-source model family for image generation and editing using DeepSeek models to interpret complex prompts and text rendering.

A new family of image models, Boogu-Image-0.1, joins the open source under an Apache 2.0 license, with the same building block for generation and editing. The Boogu project advocates this thesis from the outset: it is understanding that drives generation. Specifically, the team relies on open-source understanding models, including those from DeepSeek, to better interpret complex prompts and context.

The family covers several use cases: a Base model focused on quality and dense text, a distilled Turbo variant that generates in three or four steps with a strong photorealistic emphasis, an Edit model for instruction-based retouching, and an fp8 quantized version for constrained machines. Bilingual Chinese-English text rendering, on posters, documents, or interfaces, is among the claimed strengths.

Regarding performance, the team claims training with approximately ten times less data than open-source competitors, and a level comparable to the best closed-source systems in many cases, all on its own benchmarks, including an in-house arena set up due to lack of access to LM Arena. More unusually, Boogu also publishes a long list of acknowledged limitations, from reduced encyclopedic knowledge to still unstable editing consistency. Models, code, and demos are accessible on GitHub, Hugging Face, and via ComfyUI.