① Feature Extraction
Real Recording · ch1–5
Real Frequency Map
Raw EEG Signal
62 ch · 200 Hz · 500 ms
STFT Spectrogram
64 × 64 · freq bands
Transformer Encoder
4L · 8H · out 8×8×512
② Multi-Path Conditioning
② Semantic (Text) Path
EEG → Text Model
CLIP-like encoder
Semantic
→
Text Embedding
N × 512 · from EEG
CLIP Sim: 0.3163
③ Visual (ViT) Path
EEG → Vision Model
ViT-like encoder
Visual
→
Visual Embedding
M × 512 · spatial
④ Dynamic Conditioning
Dynamic Classifier
Cross-entropy loss
Temporal
→
Predicted Dyn. Label
Temporal sequence
58.27%
Accuracy
⑤ DANA Latent
VAE Latent
64 × 32 × 32
DANA Module
Noise injection · cond.
Adaptive
Conditioned Latent
Structured · fused
→ Fusion
⑥ SD Generation
Conditioning Fusion
Semantic + Visual + Struct.
Text Embedding (Semantic)
Visual Embedding (Visual)
DANA Latent (Structural)
SD v1.5 U-Net
Frozen + trainable attn2
Diffusion MSE
0.0005 ↓
VAE Decoder
Pretrained · fully frozen
→ pixel space
⑦ Output · ID 20
Real Observed Stimulus
Generated Video
EEG decoded · SD v1.5
SSIM
0.704
PSNR
31.44
LPIPS
0.332
— VS —
Ground Truth
Real stimulus · ID 20