Recent achievements of the team - Few-shot Image Generation via Information Transfer from the Built Geodesic Surface
Our team has published a paper titled “Few-shot Image Generation via Information Transfer from the Built Geodesic Surface” in the international journal Pattern Recognition (Impact Factor: 8.6, Top Q1 journal in the Chinese Academy of Sciences). The School of Computer Engineering and Science at Shanghai University is listed as the first affiliation.
Modern artificial intelligence models, such as generative adversarial networks and diffusion models, can generate highly photorealistic images. However, their performance remains heavily reliant on training with large-scale datasets comprising tens of thousands of images. This dependence on “big data” constitutes a significant bottleneck in numerous professional domains where collecting such vast datasets proves prohibitively costly, time-consuming, and impractical. When data is insufficient, existing models frequently fail, generating blurred or repetitive images – a predicament termed “pattern collapse”.
To tackle this challenge, our team has proposed a novel framework termed “Information Transfer from the Built Geodesic Surface (ITBGS). This approach requires no reliance on pre-trained models, instead employing an intelligent method to extract rich, deep structural information from a minimal number of samples. The framework projects image features into a high-dimensional pre-shaped space, constructs a geodesic surface, and performs sampling to generate diverse and plausible pseudo-features that simulate large-scale datasets. This augmented information is used to train an image generator, guiding it to learn complex data distribution patterns without exposure to any actual large-scale datasets.
The team’s research outcomes have been validated across multiple highly challenging datasets. Experimental results demonstrate that the ITBGS framework significantly enhances the quality, detail richness, and diversity of generated images while effectively mitigating pattern collapse issues.
Essay: Few-shot Image Generation via Information Transfer from the Built Geodesic Surface
Code: https://github.com/han-yuexing/ITBGS
