Boogu-Image-0.1 is an open-source image generation and editing model family designed to deliver near-closed-source performance using significantly less training data. It targets engineers, researchers, and power users seeking advanced multimodal capabilities, offering practical solutions for high-quality text-to-image generation, rapid image creation, and sophisticated image editing, thereby advancing the open-source AI ecosystem.

How It Works

This project introduces a unified model family, including Base, Turbo, and Edit variants, built upon systematic improvements in model understanding, data quality, and training pipelines. The approach prioritizes efficiency, achieving competitive results with an order of magnitude less training data than some comparable open-source models. The Boogu-Image-0.1-Turbo variant is a distilled model for fast, photorealistic generation, while Boogu-Image-0.1-Edit focuses on image transformation tasks.

Quick Start & Requirements

Installation involves creating a conda environment with Python 3.10, installing PyTorch (up to 2.11.0) with CUDA (up to 12.8) support, and then running pip install -r requirements/<torch>_<cuda>.txt followed by pip install -e .. A utility script utils/get_flash_attn.py assists with Flash Attention installation. Model checkpoints must be downloaded separately using huggingface-cli download into a local models/ directory. Online demos are available at http://demo-base.boogu.org/, http://demo-edit.boogu.org/, and http://demo-turbo.boogu.org/. Detailed inference options are in INFERENCE_GUIDE.md.

Highlighted Details

• Photography Generation: Accurately interprets prompts to produce high-quality images with natural lighting, coherent composition, and faithful details, even in complex scenes.
• Text Rendering: Capable of generating diverse text-heavy designs like posters and documents, featuring readable structure, stable typography, and robust bilingual (Chinese/English) rendering.
• Stylization: Offers prompt-aware creative generation across various aesthetics, including miniature 3D scenes, Chinese-inspired gilded art, fantasy visuals, and anime portraits.
• Performance: Boogu-Image-0.1 family models rank among the top-performing systems in Boogu Arena evaluations, competing with both open- and closed-source solutions.

Maintenance & Community

Recent news includes the release of ComfyUI-Boogu integration and upcoming variants like Boogu-Image-0.1-Edit-Turbo and Boogu-Image-0.1-Turbo-2K. The project maintains a presence on Hugging Face and ModelScope.

Licensing & Compatibility

This project is released under the permissive Apache-2.0 License, which generally allows for commercial use and integration into closed-source projects.

Limitations & Caveats

The model exhibits limitations in world knowledge, common sense, and understanding of real-world brands or celebrities compared to leading closed-source systems. Image-to-image consistency for editing tasks is not yet fully stable, and text rendering can falter with long text, dense typography, or complex layouts, with primary optimization for Chinese and English. Complex body structures in challenging poses and details like small faces or limbs may still show artifacts due to the VAE used. The release scope is limited due to resource constraints, and the models are intended for research, requiring additional safeguards for production deployment.