Enhancing Quantum Diffusion Models for Complex Image Generation

---

DOI: 10.13140/RG.2.2.22364.04483

---

QAMP 2025 Project > Enhancing Quantum Diffusion Models for Complex Image Generation

This repository contains the official implementation of the Hybrid Quantum-Classical U-Net, a novel architecture designed to overcome the scalability and expressibility limitations of Quantum Diffusion Models (QDMs). By integrating a Quantum Bottleneck with Adaptive Non-Local Observables (ANO) into a classical U-Net structure, this model successfully generates multi-class MNIST digits (0-9) with high structural coherence.

---

Key Features

• Hybrid Architecture: Combines the feature extraction power of classical encoders/decoders with the high-dimensional expressivity of a Quantum Bottleneck ($N=4$ qubits).
• Adaptive Non-Local Observables (ANO): Utilizes trainable Hermitian observables to extract rich, non-local features from the quantum state, solving the "measurement bottleneck" of standard VQCs.
• Skip Connections: Mitigates information loss during the quantum compression phase (256 dims $\to$ 4 qubits), ensuring image sharpness.
• Multi-Class Generation: Successfully mitigates mode collapse, generating distinct samples for all 10 digit classes (0-9), a significant improvement over prior binary-class quantum diffusion models.

Model Architecture

The model follows a U-Net design where the bottleneck layer is replaced by a Parameterized Quantum Circuit (PQC).

1. Classical Encoder: Compresses $16 \times 16$ input images into a latent vector.
2. Quantum Bottleneck: Maps latent vectors to a 4-qubit Hilbert space via Amplitude Encoding and processes them with a Variational Quantum Circuit (VQC).
3. ANO Measurement: Extracts quantum features using trainable measurements.
4. Classical Decoder: Reconstructs the image, aided by Skip Connections from the encoder.

Code Available

The source codes are in /implementation/source directory. Please check README in there.