Recently, a team from the National University of Singapore (NUS) released an innovative project called "OmniConsistency," which aims to reproduce the consistency of OpenAI's GPT-4o model in image stylization at extremely low cost. This technology not only resolves the contradiction between current open-source communities in image stylization and consistency but also provides developers with a feasible solution.
In recent years, the technology for image stylization has continued to develop. However, in practical applications, balancing style and content consistency has always been a challenge. To enhance the stylization effect, many models often sacrifice detail and semantic accuracy. The research team from NUS recognized this issue and aimed to achieve a perfect combination of stylization effects and consistency.
The core innovation of OmniConsistency lies in its unique learning framework. Unlike previous methods, OmniConsistency does not solely rely on stylization results for training; instead, it learns the consistency rules in style transfer through paired image data. Using only 2,600 high-quality image pairs and 500 hours of GPU computing power, this project achieved impressive results. Such low costs significantly reduce the burden on developers.
Additionally, OmniConsistency adopts a modular architecture that supports plug-and-play compatibility with various existing style LoRA (low-rank adaptation) modules. This means developers can easily integrate OmniConsistency into their projects without worrying about conflicts with existing systems.
Through this new technology, NUS hopes to inject nearly commercial-grade capabilities into the open-source ecosystem, providing more convenience for developers and creators. In the future, OmniConsistency may become an important tool in the field of image generation, further advancing the development of AI art creation.
Project address: https://github.com/showlab/OmniConsistency