MindPilot: Closed-loop Visual Stimulation Optimization for Brain Modulation with EEG-guided Diffusion
Dongyang Li, Kunpeng Xie, Mingyang Wu, Yiwei Kong, Jiahua Tang, Haoyang Qin, Chen Wei, Quanying Liu

TL;DR
MindPilot introduces a novel non-invasive closed-loop framework that uses EEG feedback to optimize naturalistic image generation, enabling targeted brain modulation and advancing brain-computer interface capabilities.
Contribution
It is the first framework to use EEG signals as feedback for guiding natural image generation in a closed-loop manner, without requiring explicit rewards or gradients.
Findings
Efficient retrieval of semantic targets using EEG-guided optimization
Successful closed-loop optimization of EEG features
Validation of the approach in mental matching and emotion regulation tasks
Abstract
Whereas most brain-computer interface research has focused on decoding neural signals into behavior or intent, the reverse challenge-using controlled stimuli to steer brain activity-remains far less understood, particularly in the visual domain. However, designing images that consistently elicit desired neural responses is difficult: subjective states lack clear quantitative measures, and EEG feedback is both noisy and non-differentiable. We introduce MindPilot, the first closed-loop framework that uses EEG signals as optimization feedback to guide naturalistic image generation. Unlike prior work limited to invasive settings or low-level flicker stimuli, MindPilot leverages non-invasive EEG with natural images, treating the brain as a black-box function and employing a pseudo-model guidance mechanism to iteratively refine images without requiring explicit rewards or gradients. We…
Peer Reviews
Decision·ICLR 2026 Poster
Novel and significant problem: Closed-loop visual modulation using non-invasive EEG, advancing towards "neural feedback-guided generative modeling." Practical implementation: Gradient-free black-box guidance + simple score propagation + roulette wheel sampling, resulting in low engineering cost. Multi-level validation: Proxy simulation (Tables 1, 3), dual targets of semantics and PSD (Figs. 3–5), small-scale real-time human closed-loop experiments (Fig. 6), and the provision of anonymous code.
1. Motivation and advantages of the "pseudo-model" need highlighting: The current discussion is insufficient regarding why Gaussian Process was chosen as the pseudo-model over other black-box optimizers (e.g., Bayesian Optimization). It is recommended to add a brief discussion or comparative experiment in the main text or appendix to explain the rationale for selecting GP and its advantages relative to other methods. 2. The hyperparameters α and β in Eqs. (3) and (4) are set as fixed values, b
1. The paper addresses a research problem that I believe has considerable practical value. 2. According to the authors’ evaluation, the method proposed in this paper is effective. 3. To facilitate a better evaluation, the authors conducted a visual rating experiment with human participants and demonstrated that the results obtained by their method are highly correlated with the ground truth provided by the participants. 4. The evaluation in the paper is thorough. 5. The paper is well-constru
This paper has some limitations, but I think the authors have provided a thorough discussion of them in the “Limitations” section on page 9.
Strengths: - The paper tackles a challenging and interesting problem. - The paper is well-written and aims for rigorous experimentation.
Weaknesses: - The paper claims to be the first to tackle the problem, but the principles of the approach have been extensively explored in the last years: Many core-related works are missing, which hinders the novelty statement of the papers. Examples (not exhaustive list) of a few recent references of well-known papers and some even using similar CLIP approaches, generative modeling, and diffusion processes: https://proceedings.neurips.cc/paper_files/paper/2024/hash/84bad835faaf48f24d990072
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsEEG and Brain-Computer Interfaces · Emotion and Mood Recognition · Face Recognition and Perception
