Skip to content
Help Build a Better ComfyUI Knowledge Base Become a Patron
NewsNational University of Singapore Releases OmniConsistency - Achieving Image Stylization Consistency at Low Cost

National University of Singapore Releases OmniConsistency: Achieving Image Stylization Consistency at Low Cost

OmniConsistency Demo

The Show Lab team at the National University of Singapore (NUS) recently released an open-source project called “OmniConsistency,” which can achieve image stylization consistency effects comparable to OpenAI’s GPT-4o at extremely low cost. This technology provides a practical solution for AI image generation enthusiasts and developers.

Solving Core Challenges in Image Stylization

In the field of AI image generation, balancing stylization and content consistency has always been a technical challenge. Traditional image stylization methods often face a dilemma: to achieve stronger stylistic effects, models may lose important details and semantic information from the original image.

The OmniConsistency project was born to solve this problem. This technology can maintain strong stylization effects while ensuring that generated images maintain high consistency with the original content.

Technical Features and Advantages

Innovative Learning Framework

OmniConsistency adopts a unique learning approach. Unlike traditional methods, it doesn’t solely rely on stylization results for training, but learns consistency patterns in style transfer through paired image data. This approach allows the model to better understand how to maintain content integrity during style conversion.

Extremely Low Training Cost

The most impressive feature of this project is its control over training costs. The research team used only:

  • 2,600 pairs of high-quality image data
  • 500 hours of GPU computation for training

Such training costs are significantly lower compared to other similar projects, making it affordable for more developers to develop and apply similar technologies.

Modular Design

OmniConsistency adopts a modular architecture that supports plug-and-play integration into existing systems. Particularly, its compatibility with various stylization LoRA (Low-Rank Adaptation) modules allows users to easily integrate this technology into their projects.

ComfyUI Integration Support

To make this technology more accessible to users, community developers have created dedicated node plugins for ComfyUI. Through this plugin, users can directly use OmniConsistency functionality within the ComfyUI interface.

Main Functional Features

  • Supports various LoRA modules based on FLUX.1
  • Provides multiple built-in style options, including 3D Chibi, American Cartoon, Chinese Ink painting, and 22 other styles
  • Supports custom parameter adjustments such as guidance scale and inference steps
  • Compatible with existing ComfyUI workflows

System Requirements

Note that running OmniConsistency requires high hardware specifications, with at least 40GB VRAM GPU devices recommended for optimal experience.

Rich Style Selection

OmniConsistency provides 22 different pre-trained styles, covering various fields from traditional art to modern design:

  • Traditional art styles: Oil painting, Van Gogh style, Picasso style, Chinese ink painting
  • Animation cartoon styles: Studio Ghibli, American cartoon, 3D Chibi, Snoopy
  • Modern design styles: Pixel art, vector graphics, paper craft, LEGO blocks
  • Special material styles: Fabric texture, macaron colors, origami art

Each style has been carefully trained to achieve high-quality style conversion while maintaining the original image content.

Open Source Ecosystem Contribution

By open-sourcing the OmniConsistency project, the NUS team hopes to inject more commercial-grade technical capabilities into the open-source AI community. This approach not only lowers technical barriers but also provides practical tools for more creators and developers.

The open-source nature of this project means users can:

  • Use and modify source code for free
  • Conduct secondary development based on the project
  • Share improvements and optimization solutions with the community
  • Learn advanced image stylization techniques

Future Development Prospects

With the continuous development of AI image generation technology, projects like OmniConsistency are likely to become important foundational tools in this field. It not only provides solutions for current applications but also lays the technical foundation for more innovative applications in the future.

The research team stated that they will continue to optimize algorithm performance, reduce hardware requirements, and explore more application scenarios. Active community participation and feedback will also drive continuous improvement of the project.

Through the OmniConsistency project, the National University of Singapore team has brought a practical and efficient solution to the AI image generation field. The open-source release of this technology not only advances academic research but also provides powerful tool support for developers and creators worldwide.