Brain-to-Image Generation: Reconstructing Visual Stimuli from EEG using Generative Adversarial Networks

Harshit Goyal

cs.CVSep 13, 2026

Brain-to-Image Generation: Reconstructing Visual Stimuli from EEG using Generative Adversarial Networks

Authors: Harshit Goyal

Organizations: BITS Pilani, India

Abstract

Reconstructing visual stimuli from electroencephalography (EEG) is difficult because scalp measurements have high temporal but limited spatial resolution, and paired EEG-image datasets remain small relative to modern generative-model training corpora. We present a reproducible single-subject baseline on THINGS-EEG2 that first tests the more defensible question of whether EEG can retrieve the viewed stimulus in a visual embedding space. A compact temporal-spatial convolutional encoder maps repetition-averaged EEG (63 by 250) to provided 512-dimensional ViT-B/32 image features. Model selection uses a concept-disjoint validation split, and final evaluation uses the official 200-image, 200-concept test gallery. Across three training seeds, the model obtains 12.83 +/- 0.58%, 39.17 +/- 1.76%, and 58.00 +/- 1.73% image recall at 1, 5, and 10 (mean +/- sample standard deviation), compared with analytical chance levels of 0.5%, 2.5%, and 5.0%. A session-balanced ablation shows that averaging more test repetitions generally improves ranking. Applying the Subject 01 model to the other nine subjects without adaptation causes a sharp performance drop, exposing subject specificity. We further report exploratory stress tests of direct conditional generators trained without external visual weights: single-subject and ten-subject variants produce noise-dominated outputs, with early validation improvements reversing after one to four epochs. Finally, we distinguish direct reconstruction from semantic rendering with a pretrained diffusion prior. The results support above-chance coarse semantic decoding under a closed-set, repetition-averaged protocol, but do not support faithful recovery of stimulus pixels.

Figures & tables

Partition	Images	Concepts	Repetitions/image
Optimization	14,890	1,489	4
Validation	1,650	165	4
Official test	200	200	80

Table 1: Primary Subject 01 data used in this study. Repetitions are averaged into one encoder input per image.

Figure 1: Representative stimulus images from eight distinct concepts in the Subject 01 optimization partition. The complete optimization set contains 14,890 images spanning 1,489 concepts; this montage illustrates its visual diversity rather than a cherry-picked performance subset.

Figure 2: Primary evaluation pipeline. The EEG encoder predicts a normalized visual feature, which is compared by cosine similarity with candidate image features. Image synthesis is not part of the primary retrieval result.

Stage	Operation	Output channels	Temporal width
Input	EEG tensor	63	250
Temporal	Conv2D, kernel $1\times 25$	32	250
Spatial	grouped Conv2D, kernel $63\times 1$	64	250
Pool 1	average pool $1\times 4$	64	62
Separable	depthwise $1\times 15$ , pointwise	128	62
Pool 2	average pool $1\times 4$	128	15

Table 2: EEG encoder architecture. The model has 999,008 trainable parameters; 98.6% are in the final projection.

Setting	Value	Setting	Value
Optimizer	AdamW	Batch size	128
Learning rate	$3\times 10^{-4}$	Weight decay	$10^{-3}$
Epoch budget	12	Dropout	0.35
Temperature	0.07	Cosine weight	0.25
Gradient clip	1.0	Validation fraction	0.10

Table 3: Primary retrieval training configuration.

Metric	Model	95% CI	Chance	Shuffled EEG
Recall@1	13.5%	[9.4, 18.9]	0.5%	0.5%
Recall@5	41.0%	[34.4, 47.9]	2.5%	2.0%
Recall@10	59.0%	[52.1, 65.6]	5.0%	3.5%

Table 4: Subject 01 retrieval on the official 200-concept test gallery. Confidence intervals are 95% Wilson intervals over 200 test concepts.

Figure 3: Retrieval accuracy substantially exceeds analytical chance and a fixed shuffled-EEG control at each value of $k$ .

Seed	R@1	R@5	R@10	Median rank
42	13.5%	41.0%	59.0%	8
7	12.5%	39.0%	59.0%	8
21	12.5%	37.5%	56.0%	9
Aggregate	$12.83\pm 0.58$ %	$39.17\pm 1.76$ %	$58.00\pm 1.73$ %	$8.33\pm 0.58$

Table 5: Official Subject 01 test results across independently initialized training runs. The last row reports mean

\pm

sample standard deviation.

Repetitions	R@1	R@5	R@10	Median rank	Mean rank
4	9.0%	24.5%	37.0%	19	35.87
8	15.0%	31.5%	47.5%	12	26.28
20	15.5%	36.0%	48.0%	11	20.45
40	13.5%	38.0%	56.0%	9	17.76
80	13.5%	41.0%	59.0%	8	17.11

Table 6: Session-balanced repetition ablation for the fixed seed-42 model. Each condition contributes equally from all four test sessions.

Figure 4: Retrieval versus the number of session-balanced test repetitions. Higher- $k$ recall improves overall as averaging suppresses trial noise.

Subj.	R@1	R@5	R@10	Med.	Subj.	R@1	R@5	R@10	Med.
1	13.5	41.0	59.0	8	6	4.5	11.5	19.0	39
2	4.0	16.0	24.5	28	7	1.0	10.5	18.5	42
3	2.0	10.5	18.5	48	8	3.0	13.5	20.5	40
4	4.5	14.0	22.5	31	9	2.0	8.5	16.5	46
5	2.5	9.0	15.5	40	10	5.0	20.0	33.5	21

Table 7: Subject 01 model evaluated without adaptation on all subjects. Analytical chance is 0.5%, 2.5%, and 5.0% at ranks 1, 5, and 10.

Figure 5: Zero-adaptation transfer of the Subject 01 encoder. Performance drops sharply for every other participant, indicating subject-specific neural-to-visual mappings.

Figure 6: Qualitative outputs for four held-out test stimuli. The direct feature-to-pixel decoder preserves only coarse color and layout. Both generators trained from random initialization are noise-dominated. The pretrained rendering is sharp because a large diffusion prior receives the EEG-retrieved class label; it is a semantic visualization, not recovery of the target image.

Explore similar work

Brain-to-Image Retrieval and Reconstruction via Multimodal EEG Alignment
May 18, 2026Chi Kit Wong, Yan Liu, Haowen YanEEG-Based Visual Decoding Cross-Modal Alignment
Toward High-Fidelity Visual Reconstruction: From EEG-Based Conditioned Generation to Joint-Modal Guided Rebuilding
Mar 20, 2026Zhijian Gong, Tianren Yao, Wenjia Dong +1Image Reconstruction EEG-Based Visual Decoding
Beijing University of Technology, Beijing · Beijing Key Laboratory of Computational Intelligence and Intelligent System, Beijing
EEG2Vision: A Multimodal EEG-Based Framework for 2D Visual Reconstruction in Cognitive Neuroscience
Apr 9, 2026Emanuele Balloni, Emanuele Frontoni, Chiara Matti +3EEG-Based Visual Decoding EEG Decoding
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 +1

May 18, 2026J/K move · Enter open · S save

Brain-to-Image Retrieval and Reconstruction via Multimodal EEG Alignment

Chi Kit Wong, Yan Liu, Haowen Yan

EEG-Based Visual Decoding Cross-Modal Alignment Image Reconstruction Cross-Modal Retrieval

Open paper AskPDF arXiv

We present a brain-to-image system that decodes visual stimuli from EEG signals recorded during natural image viewing. Our system addresses two tasks: (1) EEG-to-image retrieval, which ranks the correct stimulus image among 200 candidates given an EEG segment, and (2) EEG-to-image reconstruction, which generates an image consistent with the perceived stimulus. For retrieval, we implement a multi-level blurring approach improved with biologically inspired EVNet features and trained with the InfoNCE loss. Evaluated over 10 random seeds for a single subject, the retrieval model achieves a mean final-epoch Top-1 accuracy of 86.30% and Top-5 accuracy of 98.55%. For reconstruction, we implement CognitionCapturerPro, which aligns EEG representations to multi-modal CLIP embeddings, including image, text, depth, and edge embeddings, and synthesizes images with SDXL-Turbo conditioned via IP-Adapter. Averaged over 10 seeds, the reconstruction model achieves a CLIP score of 0.903 using ViT-H-14, a CLIP score of 0.870 using ViT-L/14, and an SSIM of 0.409. These results demonstrate the feasibility of decoding rich visual representations from EEG signals using modern multi-modal alignment and generative modeling techniques.

Abstract

Explore similar work

Brain-to-Image Retrieval and Reconstruction via Multimodal EEG Alignment

Toward High-Fidelity Visual Reconstruction: From EEG-Based Conditioned Generation to Joint-Modal Guided Rebuilding

EEG2Vision: A Multimodal EEG-Based Framework for 2D Visual Reconstruction in Cognitive Neuroscience