SPSC reconstructs images from unseen subjects' EEG without labels, lifting CLIP from 0.552 to 0.713

Cross-subject EEG reconstruction can be semantically corrected without target labels or target image embeddings, with SPSC applying the correction after the EEG projector and editing only the semantic condition fed to the diffusion prior and generator. Earlier EEG-to-image methods mainly optimized one-to-one trial-image alignment, so a projector trained on source subjects produced target embeddings whose global distribution and local coordinates shifted away from the image-embedding manifold. Under leave-one-subject-out evaluation across ten subjects on THINGS-EEG, the SPSC reconstruction readout raises CLIP from 0.552 to 0.713 and Alex.5 from 0.612 to 0.745 against the Li et al. method.

The training-side object-mask contrastive regularizer treats different images of the same object concept as positives, keeping EEG responses to one concept adjacent in the conditioning space. With paired alignment alone, different images under one object concept were treated as unrelated negatives, leaving the conditioning space at discrete matched pairs. In the cross-subject Sub-08 ablation, adding this regularizer raises CLIP from 0.527 to 0.679, and enabling the reconstruction readout's semantic correction on top of it raises CLIP further to 0.723.

The inference-side subspace candidate gate builds candidate semantics in PCA subspaces of image prototypes, aligns distributions with unlabeled target EEG batch statistics, and routes each sample by cosine confidence against training prototypes. The correction rule and hyperparameters are fitted only on source subjects, while the target side contributes only unlabeled EEG batch statistics; stimulus identities and image embeddings are used for evaluation only after inference. On the Sub-08 test, the confidence selector reaches 24.0 Top-1 on the retrieval readout, against 23.5 for fixed k256, 21.0 for candidate averaging, and 27.0 for a post-hoc oracle that picks by test rank.

Evaluation covers only the THINGS-EEG dataset under a ten-subject leave-one-subject-out protocol, the reconstruction pipeline uses SDXL-Turbo with IP-Adapter, and all experiments run on a single 24 GB RTX 4090. Target EEG arrives in batches, inference uses only its unlabeled statistics, and stimulus identities and image embeddings are used for evaluation only after inference.

The source lists THINGS-EEG as the only publicly available dataset for this work, so the conclusions are scoped to that dataset and these subjects. Qualitative results still include errors ranging from complete category loss to keeping a coarse cue such as clothing or transportation while losing object identity, and furniture and electronic devices show weak recovery of target-specific appearance and geometry. A gap remains between confidence selection and the post-hoc oracle that picks by test rank.