Sabi enables AI to learn personal aesthetics from brain signals, generating faces that increasingly match preferences.

Beating AI News Flash: Brain-computer interface company Sabi has published new research called RLNF (Reinforcement Learning from Neural Feedback). In the demo, an AI-generated face continuously changes, gradually becoming more aligned with the viewer's aesthetic preferences. The researchers hope to let AI learn each person's preferences from brain signals, without having to repeatedly ask users which image looks better.

The team first had 5 subjects view a large number of faces while recording EEG and eye movements, and asked them to rate the attractiveness of each face. From this data, the AI learned what physiological responses each person shows when seeing faces they like or dislike, thereby building a personalized aesthetic preference model.

Subsequently, the researchers used these preference data to fine-tune the FLUX.2 Klein 4B image model through reinforcement learning. Each time the AI generated a new batch of faces, the system compared them with previously recorded faces and gave rewards based on personal preferences inferred from brain signals. The model continuously adjusted according to the rewards and gradually learned to generate faces more aligned with a specific user's aesthetics, without requiring subjects to rate each image again.

Previous research has already used EEG feedback to guide AI in generating faces people like. Sabi instead applied this feedback to reinforcement learning of an image model, allowing personal aesthetics to directly influence the model's generation tendencies.

However, the experiment currently has only 5 participants, and the preferences of 2 of them were mainly judged through eye activity rather than directly identified from brain activity. More experiments are needed to prove that EEG can reliably reflect personal aesthetics.