NEvo: Evolutionary AI Synthesizes Videos That Maximally Activate a Chosen Brain Region

Researchers at EPFL (Amir Zamir’s lab) built NEvo, a framework that generates video clips engineered to drive a specific region of the visual cortex as hard as possible. The pipeline first trains an encoding model — a ‘digital twin’ of the brain — that predicts voxel-level fMRI responses to any video. It then runs an evolutionary search over a structured prompt space, where each candidate clip is described by parameters like object, lighting, motion, and mood. NEvo generates a batch of videos, scores each one against the digital twin’s predicted activation for the target ROI, and iteratively selects, mixes, and mutates the strongest candidates until it converges on maximally activating ‘hyper-activating’ stimuli.

The synthesized clips consistently outperform the handcrafted localizer videos neuroscientists traditionally use, and they recover known regional selectivities without being told about them: faces for FFA, places for PPA, bodies for EBA, coherent motion for MT, low-level patterns for V1/V3A, and lively social scenes for the superior temporal sulcus (pSTS/aSTS). Because the stimuli are dynamic rather than static, the method also exposes systematic differences in how sensitive each region is to temporal information, showing that selectivity grows more motion-driven and social as you move along the lateral visual stream.

The significance is methodological: this is the first model-based approach to probing visual selectivity under naturalistic, dynamic stimulation, extending earlier image-only work (XDREAM, NeuroGen, and related closed-loop fMRI efforts) into video. Rather than testing a fixed battery of hand-picked stimuli, scientists can now ask a generative model to invent the optimal stimulus for any brain area — a faster, more expressive way to map what different parts of the visual system actually care about.