Sabi enables AI to learn personal aesthetic preferences from brain signals, with the generated faces growing increasingly aligned with users’ tastes.

Beating AI News (Dongcha) — Brain-computer interface firm Sabi has unveiled a new study titled RLNF (Reinforcement Learning with Neural Feedback). In the demo, an AI-generated face evolves continuously, gradually transforming into a shape more aligned with the viewer’s aesthetic preferences. Researchers aim to enable AI to learn individual preferences from brain signals, eliminating the need to repeatedly ask users which images they prefer. The team first had 5 participants view a large number of faces, while recording their electroencephalogram (EEG) and eye movements, and asking them to rate each face’s attractiveness. Using this data, the AI learns the physiological responses each person exhibits when viewing faces they like or dislike, building a model of individual aesthetic preferences. Subsequently, researchers used the preference data to fine-tune the FLUX.2 Klein 4B image model via reinforcement learning. Each time the AI generates a new batch of faces, the system compares them to previously recorded faces and delivers rewards based on individual preferences inferred from brain signals. The model adjusts continuously based on these rewards, gradually learning to generate faces that better match a specific user’s aesthetic, with no need for participants to rate each image individually throughout the process. Previous studies have used EEG feedback to guide AI in generating faces people like. Sabi, however, applies this feedback to reinforcement learning for image models, allowing individual aesthetics to directly influence the model’s generation tendencies. That said, the experiment currently only has 5 participants, with the preferences of 2 of them being judged primarily via eye movements rather than direct identification from brain activity. More experiments are needed to confirm that EEG can reliably reflect individual aesthetic preferences.