Pacific-Prime/diffusion-DiT

ComplexityDiT - Diffusion Transformer with INL Dynamics

Diffusion Transformer enhanced with PID-style dynamics control for smoother denoising.

Architecture

Input -> [Attention -> MLP -> Dynamics] x 12 -> Output

Core equations:

• Attention: softmax(QK^T/sqrt(d)) * V
• MLP: W2 * GELU(W1 * x)
• Dynamics: h += dt * gate * (alpha*v - beta*(h - mu))

Model Details

Parameter	Value
Architecture	ComplexityDiT-S
Parameters	114M
Layers	12
Hidden dim	384
Heads	6
Experts	4
Dynamics	Enabled

Training

• Dataset: huggan/wikiart
• Steps: 20,000
• Batch size: 16
• Mixed precision: FP16

Usage

from safetensors.torch import load_file
from complexity_diffusion import ComplexityDiT

# Load model
model = ComplexityDiT.from_config('S', context_dim=768)
state_dict = load_file('model.safetensors')
model.load_state_dict(state_dict)

INL Dynamics

The dynamics layer adds robotics-grade control to stabilize denoising trajectories:

• mu - learnable equilibrium (target position)
• alpha - inertia (momentum)
• beta - correction strength (spring constant)
• gate - amplitude control

This creates smooth, stable trajectories like a PID controller guiding the model toward clean images.

Links

• GitHub
• PyPI