Organizations: Department of Civil Engineering Building and Architecture (DICEA), Universit`a Politecnica delle Marche, Via Brecce Bianche 12, Ancona, 60131, Italy · Department of Political Sciences, Communication and International Relations, University of Macerata, Via Don Minzoni 22A, Macerata, 62100, Italy · Horizon Intelligence Labs · Harvard Medical School (HMS), Precision Neuromodulation Program & Network Control Laboratory, Gordon Center for Medical Imaging, Department of Radiology, Massachusetts General Hospital, Horizon Intelligence Labs, Boston, MA, USA
Reconstructing visual stimuli from non-invasive electroencephalography (EEG) remains challenging due to its low spatial resolution and high noise, particularly under realistic low-density electrode configurations. To address this, we present EEG2Vision, a modular, end-to-end EEG-to-image framework that systematically evaluates reconstruction performance across different EEG resolutions (128, 64, 32, and 24 channels) and enhances visual quality through a prompt-guided post-reconstruction boosting mechanism. Starting from EEG-conditioned diffusion reconstruction, the boosting stage uses a multimodal large language model to extract semantic descriptions and leverages image-to-image diffusion to refine geometry and perceptual coherence while preserving EEG-grounded structure. Our experiments show that semantic decoding accuracy degrades significantly with channel reduction (e.g., 50-way Top-1 Acc from 89% to 38%), while reconstruction quality slight decreases (e.g., FID from 76.77 to 80.51). The proposed boosting consistently improves perceptual metrics across all configurations, achieving up to 9.71% IS gains in low-channel settings. A user study confirms the clear perceptual preference for boosted reconstructions. The proposed approach significantly boosts the feasibility of real-time brain-2-image applications using low-resolution EEG devices, potentially unlocking this type of applications outside laboratory settings.
Figures & tables
Figure 1: Overview of the EEG2Vision framework. The visual stimulus is encoded into a noisy latent via the Image processing module and also fed into the Semantic prior generation module, which utilizes both visual and EEG signals to derive coarse class-level controls. Simultaneously, EEG signals undergo Neural embedding extraction to produce fine-grained spatiotemporal features. These inputs condition the Diffusion-based reconstruction module through ControlNet to synthesize the generated image. Finally, the Prompt-guided post-reconstruction boosting module refines the output via text-guided image-to-image diffusion.
System Prompt : ”You are an expert in textual description from a single image. Given an image, you will provide a concise and accurate description of the content, without saying ‘the image shows’ or ‘the image depicts’ at the start.” User Prompt : ”Write a description for this image in one sentence. You should answer with the prompt only. Do not insert the first part where you say ‘the image shows’ or ‘the image depicts’ in the answer.”
Table 1: Prompt template used for the MLLM to generate concise image descriptions.
pref= ”A realistic, high-quality photo of a [d] , with clean and correct geometry, natural lighting, consistent textures, and accurate proportions. No visual glitches, no distorted shapes, no rendering artifacts. The object appears physically plausible and professionally photographed, with all structures logically and realistically aligned.”
Table 2: Prompt template used for the image-to-image diffusion model, conditioned on the image description.
Channels
CFG
50-way Accuracy
IS ↑
FID ↓
LPIPS ↓
CLIP-Sim ↑
Top-1 ↑
Top-5 ↑
128
4
0.876
0.928
33.93
79.14
0.770
0.723
7.5
0.890
0.935
34.82
76.77
0.770
0.733
64
4
0.800
0.869
33.45
83.05
0.773
0.700
7.5
0.823
0.877
34.11
78.27
0.777
0.714
32
4
0.375
0.458
33.26
86.00
0.785
0.620
Table 3: Quantitative results of different channel configurations and CFGs across multiple performance metrics.
Figure 2: Qualitative results of the evaluation for different channel resolutions and CFG configurations. Some failure cases are also shown for low accuracy classes (highlighted in red).
128 Channels
64 Channels
32 Channels
24 Channels
Metric
Raw
Boosted
Gain
Raw
Boosted
Gain
Raw
Boosted
Gain
Raw
Boosted
Gain
IS ↑
34.82
37.23
+6.69%
34.11
37.37
+9.12%
33.28
36.51
+9.71%
34.24
37.16
+8.53%
FID ↓
76.77
77.06
+0.38%
78.27
77.75
+0.67%
80.55
79.26
+1.60%
80.51
79.77
+0.92%
LPIPS ↓
0.770
0.769
+0.13%
0.777
0.773
+0.52%
0.791
0.787
+0.51%
0.790
0.785
+0.63%
Table 4: Quantitative evaluation of raw and boosted reconstructions across different electrode configurations. Percentage gains indicate improvement of the boosted method over the original.
Figure 3: Qualitative comparison of reconstructions across electrode configurations. Columns from left to right show: GT reference image, raw generated reconstruction and our refined output (boosted).
Figure 8
Channels
Boosted Preference Rate
Mean Confidence
Weighted Preference Rate
128
71.43%
3.7
74.17%
64
82.35%
3.5
85.35%
32
76.97%
3.6
79%
24
82.73%
3.6
87.13%
Overall
78.34%
3.6
81.31%
Table 5: User study results showing preference for boosted images across different channel configurations. Confidence ratings are on a 5-point Likert scale.
Department of Automation Tsinghua University Haidian District, 100084, Beijing, China · China Academy of Information and Communications Technology Haidian District, 100084, Beijing, China