BCIwiki (bciwiki.com) — According to arXiv, a preprint uploaded to the platform on 2026-10-09 introduces Neurons++, a framework for reconstructing static images and dynamic videos from functional magnetic resonance imaging (fMRI) recordings. The preprint has not been peer reviewed. The study asks not only whether generated visuals look plausible, but also whether they are grounded in brain activity, a question the authors connect to understanding human perception and developing brain-computer interfaces (BCIs).
Compared with the earlier conference version, Neurons, which addressed video reconstruction only, Neurons++ extends reconstruction to both images and videos. It replaces parallel task conditioning with serial cortex-inspired decoding. Image-fMRI pretraining first establishes a spatial-semantic initialization, followed by subject-specific fine-tuning for image and video reconstruction separately. The serial design organizes fMRI-derived representations into spatial, semantic, language-level, and temporal cues.
The authors report state-of-the-art video reconstruction and competitive image reconstruction across standard benchmarks. Their neural-to-visual analyses indicate qualitative directional controllability under neural perturbation, sensitivity of the fMRI encoder read-out to stimulus-relevant cortical features, and semantic modulation guided by regions of interest (ROIs). These findings describe model responses in the reported tests rather than comprehensively establishing faithful reconstruction of visual experience. The paper states that code and model weights are publicly available.