NEUROSKETCH — EEG & Concept → Abstract Image

NEUROSKETCH converts simulated EEG-style brainwave signals or conceptual text into abstract images. It maps simple EEG band-power features to visual adjectives (mood, color palette, texture), constructs a creative text prompt, and uses a text-to-image model (Stable Diffusion via diffusers) to generate artwork.

⚠️ Safety / scope: This repository is a creative research prototype. It is not a medical tool and should not be used with real patient EEG data in public contexts or for clinical decision-making. Use synthetic or anonymized data and follow all privacy/regulatory rules.

Features

Generate synthetic EEG data for demos (synth_eeg.py)
Map EEG rhythms (delta/alpha/beta/gamma) to mood/color/texture adjectives (eeg_to_prompt.py)
Compose artistically-rich prompts and generate images with diffusers (generate_image.py)
Streamlit app to try EEG→image or prompt→image interactively (app.py)
Notebook demo and CLI examples included

Quickstart

Create and activate a virtual environment:

python -m venv venv
source venv/bin/activate     # Windows: venv\Scripts\activate
pip install -r requirements.txt

Generate synthetic EEG:

python synth_eeg.py

Convert EEG to prompt:

python eeg_to_prompt.py --eeg examples/sample_eeg.csv

Generate an image (requires model download; may be large):

python generate_image.py --prompt "a dreamy abstract painting, soft pastels" --out_dir outputs

Or run the web demo:

streamlit run app.py

How it works (high-level)

EEG (simulated): synth_eeg.py creates a short multichannel CSV (time + channels).
Feature extraction: eeg_to_prompt.py computes band powers (low/alpha/beta/gamma) using Welch PSD.
Mapping: Heuristics map relative band power to descriptive adjectives (mood, color, texture).
Prompt assembly: Template fills adjectives into an artistic prompt.
Generation: generate_image.py uses a diffusion model (via diffusers) to produce the image.

Limitations & Ethics

Mapping EEG → imagery is speculative and artistic; not scientifically validated.
Do not use real patient EEG in public repositories without consent and compliance.
Image generation inherits biases from pretrained models — inspect outputs critically.

Downloads last month: 3