boogu-project/Boogu-Image

What it solves

Boogu-Image-0.1 is an open-source unified image generation and editing model family designed to provide high-quality text-to-image generation, fast inference, and precise image editing. It specifically addresses the challenge of achieving competitive performance in photography, stylization, and bilingual (Chinese-English) text rendering while using significantly less training compute—roughly one order of magnitude smaller data scale—than closed-source multimodal systems.

How it works

The project provides a suite of models (Base, Turbo, Edit, and Edit-Turbo) that leverage systematic improvements in understanding ability, data quality, and training pipelines.

• Base: A foundation model optimized for diversity, controllability, and ultra-dense text rendering.
• Turbo: A distilled variant using Decoupled DMD that enables high-quality photorealistic generation in only 3-4 steps.
• Edit: A variant specialized for image-to-image transformations, supporting resolutions up to 2K.
• Edit-Turbo: A fast, distilled version of the editing model.

Who it’s for

This project is for researchers and developers in the multimodal generation space who need a high-performance, open-source alternative for generating photorealistic images, designing text-heavy layouts (like posters and brand guides), or performing complex image editing tasks.

Highlights

• Bilingual Text Rendering: Stable and readable Chinese and English typography across diverse layouts.
• Precise Image Editing: Supports object insertion, replacement, removal, and material modification while maintaining subject coherence.
• High-Efficiency Inference: The Turbo variants provide photorealistic results in just 4 steps.
• Versatile Output: Supports multiple aspect ratios and resolutions up to 2K (for the Base and Edit models).
• Broad Hardware Support: Includes native PyTorch support and a dedicated branch for Ascend NPU inference.