ZEBRA: Towards Zero-Shot Cross-Subject Generalization for Universal Brain Visual Decoding

Haonan Wang (The Hong Kong University of Science and Technology) · Jingyu Lu (Peking University) · Hongrui Li (The Hong Kong University of Science and Technology) · Xiaomeng Li (The Hong Kong University of Science and Technology)

adversarial trainingbrain visual decodingdisentanglement fine-tuningfmri-to-image reconstructiongeneralization neural decoding neural representations performance metrics scalabilitysemantic-specific representationssubject-invariant representationssubject-specific modelsuniversal neural decodingvisual experienceszero-shot learning

Recent advances in neural decoding have enabled the reconstruction of visual experiences from brain activity, positioning fMRI-to-image reconstruction as a promising bridge between neuroscience and computer vision. However, current methods predominantly rely on subject-speciﬁc models or require subject-speciﬁc ﬁne-tuning, limiting their scalability and real-world applicability. In this work, we introduce ZEBRA, the ﬁrst zero-shot brain visual decoding framework that eliminates the need for subject-speciﬁc adaptation. Z EBRA is built on the key insight that fMRI representations can be decomposed into subject-related and semantic-related components. By leveraging adversarial training, our method explicitly disentangles these components to isolate subject-invariant, semantic-speciﬁc representations. This disentanglement allows ZEBRA to generalize to unseen subjects without any additional fMRI data or retraining. Extensive experiments show that ZEBRA signiﬁcantly outperforms zero-shot baselines and achieves performance comparable to fully ﬁnetuned models on several metrics. Our work represents a scalable and practical step toward universal neural decoding. Code and model weights are available at: https://github.com/xmed-lab/ZEBRA.