An EEG image decoder learns to adjust to unfamiliar people

The challenge of a new participant

Turning EEG recordings into images requires more than a powerful image generator. The system must translate noisy electrical recordings into a representation that tells the generator something meaningful about what a person saw. Differences between people—including anatomy, physiology and recording conditions—can disrupt that translation.

In a November 2026 preprint, Yutang Lu, Shuai Jiang, Fei Su and Zhicheng Zhao of Beijing University of Posts and Telecommunications describe Semantic Preservation and Subspace Calibration, or SPSC. Their approach addresses two problems: keeping related objects close together in the model’s learned representation, and adjusting that representation for an unfamiliar participant.

The experiments concern images people viewed, not freely imagined scenes or arbitrary thoughts. The findings are reported in a preprint and should be treated as preliminary research.

Keeping object relationships intact

The pipeline first converts EEG into a numerical representation of visual meaning. Further components predict an image embedding and lower-level visual features, which guide an SDXL-Turbo generator equipped with IP-Adapter.

The first addition, Object Mask Contrastive Regularization, changes how the reconstruction components learn. Ordinary paired training emphasizes matching each EEG response to its corresponding image. But different pictures of the same object concept can then be treated as unrelated alternatives.

SPSC uses training-set object labels to encourage features belonging to the same concept to stay close together. It also retains a constraint tying predictions to their individual image targets. The aim is to preserve shared object identity without abandoning the particular training image.

This regularization applies to both the predicted image embedding and the lower-level branch, although the authors describe the latter as an auxiliary constraint because those features also carry image-specific layout and texture.

Calibration without image labels

The second addition, the Subspace Candidate Gate, operates after the EEG projector. It uses representations of training images as reference points and measures the distribution of an unlabeled batch of EEG representations from the new participant.

The method constructs several possible corrections in spaces derived from the training-image representations using principal component analysis. These spaces retain different numbers of dimensions. A confidence rule then selects a complete candidate representation for each sample, based on its similarity to the training-image reference bank.

Importantly, “zero-shot” does not mean the system sees no data from the new person. It uses their unlabeled EEG batch to estimate calibration statistics. Their stimulus identities and image embeddings are reserved for evaluation, while calibration learning and parameter tuning rely on the source participants.

The approach also distinguishes two tasks. Reconstructing an image benefits from lower-dimensional, smoother conditions. Retrieving the correct image from a set of alternatives benefits from retaining more discriminating dimensions.

What the benchmark showed

The researchers evaluated SPSC on THINGS-EEG using ten participants. They repeatedly held out one participant and trained using the others. The dataset separates 1,654 training object concepts from 200 test concepts.

Across these held-out-participant tests, the reported mean CLIP reconstruction score rose from 0.552 for the comparison method by Li and colleagues to 0.713 for SPSC. The AlexNet layer-five score rose from 0.612 to 0.745.

These are two-way identification measures based on image-model features—not percentages of perfectly reconstructed pictures. They assess whether a reconstruction better matches its corresponding image than an alternative in the relevant feature space.

In a separate 200-way retrieval test, SPSC’s retrieval readout placed the correct image first in 23.7% of cases on average, compared with 16.4% for NeuroBridge, the strongest baseline reported for that measure.

Component tests on participant 08 attributed the largest reconstruction improvement to object-level training regularization, with calibration providing a further gain.

Plausible images are not necessarily faithful

The evidence comes from one dataset and ten participants; several detailed diagnostic analyses focus on participant 08. Broader transfer across people and recording sessions remains to be established.

The paper also documents failures ranging from losing the object category entirely to retaining only a broad cue. Furniture and electronic devices exposed difficulties recovering specific appearance and geometry.

That distinction matters: the generator can produce a coherent picture even when the EEG-derived condition points to the wrong object. SPSC improves a measured interface between neural recordings and image generation, but it does not establish reliable, exact recovery of what an unfamiliar person sees.