[Paper Review] ENIGMA: EEG-to-Image in 15 Minutes Using Less Than 1% of the Parameters

Reese Kneeland, Wangshu Jiang+4 more|arXiv (Cornell University)|Feb 10, 2026

EEG and Brain-Computer Interfaces0 citations

TL;DR

ENIGMA is a multi-subject EEG-to-image model that fine-tunes on new subjects in as little as 15 minutes, uses <1% of the parameters, and achieves state-of-the-art reconstructions across THINGS-EEG2 and Alljoined-1.6M, with robust performance on consumer hardware.

ABSTRACT

To be practical for real-life applications, models for brain-computer interfaces must be easily and quickly deployable on new subjects, effective on affordable scanning hardware, and small enough to run locally on accessible computing resources. To directly address these current limitations, we introduce ENIGMA, a multi-subject electroencephalography (EEG)-to-Image decoding model that reconstructs seen images from EEG recordings and achieves state-of-the-art (SOTA) performance on the research-grade THINGS-EEG2 and consumer-grade AllJoined-1.6M benchmarks, while fine-tuning effectively on new subjects with as little as 15 minutes of data. ENIGMA boasts a simpler architecture and requires less than 1% of the trainable parameters necessary for previous approaches. Our approach integrates a subject-unified spatio-temporal backbone along with a set of multi-subject latent alignment layers and an MLP projector to map raw EEG signals to a rich visual latent space. We evaluate our approach using a broad suite of image reconstruction metrics that have been standardized in the adjacent field of fMRI-to-Image research, and we describe the first EEG-to-Image study to conduct extensive behavioral evaluations of our reconstructions using human raters. Our simple and robust architecture provides a significant performance boost across both research-grade and consumer-grade EEG hardware, and a substantial improvement in fine-tuning efficiency and inference cost. Finally, we provide extensive ablations to determine the architectural choices most responsible for our performance gains in both single and multi-subject cases across multiple benchmark datasets. Collectively, our work provides a substantial step towards the development of practical brain-computer interface applications.

Motivation & Objective

• Address practical EEG-to-Image decoding for real-world BCI by enabling rapid fine-tuning on new subjects.
• Achieve robust performance on both research-grade and consumer-grade EEG hardware.
• Reduce model size by sharing parameters across subjects while preserving decoding quality.
• Provide comprehensive evaluation including human behavioral assessments and ablations.
• Demonstrate broad applicability and efficiency for edge deployment and clinical use.

Proposed method

• Propose ENIGMA, a multi-subject EEG-to-image model with a spatio-temporal backbone, subject-wise latent alignment layers, and an MLP projector to a CLIP embedding space.
• Use a unified multi-subject architecture with lightweight subject-specific alignment to share most parameters across subjects.
• Map EEG embeddings to CLIP ViT-H/14 latent space and reconstruct images via Stable Diffusion XL Turbo with an IP-Adapter.
• Train with a composite loss combining MSE between EEG embeddings and image CLIP embeddings and an InfoNCE contrastive term.
• Allow three operation modes: single-subject, multi-subject, and fine-tuned subject adaptation.
• Demonstrate 15-minute calibration feasibility and potential edge-device deployment; report training efficiency (e.g., 5.5 hours for 30 subjects).

Experimental results

Research questions

• RQ1 Can a unified multi-subject EEG-to-image model achieve state-of-the-art reconstruction across both high-quality and consumer-grade EEG hardware?
• RQ2 Does a lightweight, shared-parameter architecture with subject-specific latent alignment enable rapid adaptation to new subjects with minimal data?
• RQ3 How does ENIGMA perform relative to existing EEG-to-Image baselines on standard benchmarks (THINGS-EEG2 and Alljoined-1.6M) in both automated and human evaluations?
• RQ4 What is the impact of architectural components (latent alignment, spatio-temporal backbone, diffusion priors) on cross-subject generalization and robustness to hardware quality?
• RQ5 Is ENIGMA scalable to multiple subjects with substantial parameter efficiency compared to single-subject models?

Key findings

• ENIGMA achieves SOTA performance on THINGS-EEG2 and Alljoined-1.6M across multiple metrics and exhibits robust cross-subject generalization via latent alignment.
• The model uses <1% of the trainable parameters required by prior approaches and scales to 30 subjects with a ~165x parameter reduction for multi-subject deployments.
• ENIGMA enables fine-tuning on a new subject with as little as 15 minutes of data, outperforming non-pretrained baselines in low-data regimes.
• Human behavioral evaluations show ENIGMA reconstructions are more identifiable to ground truth images than baselines across conditions.
• Ablation analyses show latent alignment and the spatio-temporal backbone are critical for multi-subject performance, while certain diffusion-prior components may hurt performance on consumer hardware.
• Across benchmarks, ENIGMA maintains robust performance on consumer-grade EEG hardware, reducing brittleness observed in more complex architectures.