This project provides an enterprise-level AI image generation platform, BaiduCBIT, built on ComfyUI, specializing in photorealistic human image synthesis. It targets users requiring high-fidelity visual content, offering an end-to-end solution from text prompts to detailed, realistic images through its advanced Flux model architecture and integrated AI enhancement technologies. The platform aims to streamline complex generation workflows with a user-friendly interface and robust backend.

How It Works

BaiduCBIT leverages the Flux 1.0 Dev Model, a 12B parameter Diffusion Transformer architecture, enhanced with FP8_E4M3FN quantization for memory efficiency. It employs a Dual-CLIP architecture for robust text understanding and integrates specialized LoRA modules for fine-tuning details like hands and realism. Generation utilizes a dual-stage sampling strategy with adaptive guidance and ControlNet for precise control. A multi-level post-processing pipeline refines output for photographic quality. The system supports both production-ready distributed deployment and local development environments.

Quick Start & Requirements

• Installation: Clone the repository, install dependencies (pip install -r requirements.txt), configure environment variables (cp env.example .env), and start the service (python run.py).
• Prerequisites: Linux (Ubuntu 20.04+ recommended), Python 3.12+, NVIDIA GPU (RTX 3080+ with 8GB+ VRAM), CUDA 12.4+, 16GB+ RAM, 100GB+ storage.
• Links: GitHub Repository.

Highlighted Details

• Achieves photorealistic human image generation with detailed skin, facial features, hair, and lighting.
• Includes specialized LoRA modules for hand repair and realism enhancement.
• Supports multilingual prompt translation via Bing Translator API.
• Offers precise control over generation using various ControlNet types (Tile, Depth, Pose, Edge).
• Optimized for performance with FP8 quantization, achieving ~15-30 second generation times on an RTX 4090 for 30 steps.

Maintenance & Community

• Maintainer: BaiduCBIT Team.
• Developer: @reneverland.
• Support: Primarily through GitHub Issues.
• Roadmap: Features in development include video generation and batch processing optimization. Planned features include support for SDXL/SD3.5, real-time preview, and user permission management.

Licensing & Compatibility

• License: MIT License.
• Compatibility: The MIT license is permissive, generally allowing commercial use and integration into closed-source projects.

Limitations & Caveats

The project is actively under development, with key features like video generation and broader model support (SDXL, SD3.5) still in the planning or development stages. A specific CUDA version (12.4+) is required. The demo video is not included in the repository.