alindumitru/neural-art-eeg

Neural Art EEG-to-Image Model

This repository contains a custom PyTorch latent diffusion checkpoint trained with the neural-art repository code. It generates images conditioned on Emotiv Epoch X style EEG input.

Learn more about Neural Art

Github: https://github.com/alinvdu/neural-art Youtube: https://www.youtube.com/watch?v=8v_EB73m6cQ

Files

• generative/checkpoints/checkpoint.pth: fine-tuned latent diffusion checkpoint.
• generative/checkpoints/checkpoint-eeg.pth: EEG encoder checkpoint required by the current inference code.
• config15.yaml: model architecture configuration.
• generative/: custom model code copied from src/generation-service/generative.
• inference.py: small helper for loading the model from this Hub repository.

Usage

This is custom PyTorch code, not a standard transformers or diffusers pipeline. Install the runtime dependencies, clone this repository, and run inference from the repository root.

pip install -r requirements.txt

import torch
from inference import NeuralArtPipeline

pipe = NeuralArtPipeline(device="cuda" if torch.cuda.is_available() else "cpu")

# EEG input should match the training/inference shape used by the original repo:
# a 2D tensor/array with shape [channels, time].
eeg = torch.randn(14, 1024)
images = pipe(eeg, num_samples=2, ddim_steps=100)

for i, image in enumerate(images):
    image.save(f"sample_{i}.png")

Intended Use

Research and experimentation with EEG-conditioned image generation. This model was developed for data compatible with an Emotiv Epoch X 14-channel headset and the preprocessing pipeline in the original repository.

Limitations

• The model depends on custom repository code.
• Inputs outside the training distribution may produce poor or unstable results.
• The current loader expects both the diffusion checkpoint and the EEG encoder checkpoint.

Citation / Credits

Original repository: neural-art

Architecture credit in the source repository: DreamDiffusion.