AI-Generated Videos That Evolve to Maximally Activate Target Brain Regions

Proposes the NEvo methodology, which builds a digital twin (Encoding Model) that predicts the brain's visual responses and uses it as a reward function to find videos that most powerfully stimulate specific brain regions.

How it works:

• Evolutionary prompt optimization: Generates a population of prompts with 'genes' such as subject, lighting, movement, and mood, then finds the optimal video through crossover and mutation based on the digital twin's prediction scores.
• Two-stage generation process: First finds the most powerful still image, then adds motion to that image to complete a 2-second clip.

Key results:

• High activation rate: Videos generated by NEvo activate the target brain region more strongly than existing hand-crafted clips or natural videos.
• Mapping of social features: Confirmed that as you move from the visual cortex (V1) to the anterior superior temporal sulcus (aSTS), the generated videos evolve from simple patterns to forms that include faces and social interactions between people.
• Feature isolation: Even starting from abstract stimuli, the optimization process can clearly extract the features preferred by specific regions (e.g., motion for the MT area, social interaction for the pSTS area).