"PassionSR: Post-Training Quantization with Adaptive Scale in One-Step Diffusion based Image Super-Resolution"

Smart quantization meets super-resolution: PassionSR makes AI models hardware-friendly

PassionSR squeezes one-step diffusion models into tiny packages without losing their super-resolution magic

PassionSR introduced in this paper, is a novel post-training quantization approach for one-step diffusion models in image super-resolution. It tackles the challenge of high computational and storage costs by simplifying the model architecture and introducing adaptive quantization techniques, making deployment on hardware devices feasible.

-----

https://arxiv.org/abs/2411.17106v1

🔍 Original Problem:

One-step diffusion models for image super-resolution, while faster than multi-step approaches, still demand significant computational resources and storage, limiting their practical deployment on hardware devices.

-----

🛠️ Solution in this Paper:

→ The model architecture is simplified to two core components: UNet and Variational Autoencoder (VAE), removing the CLIPEncoder.

→ A Learnable Boundary Quantizer (LBQ) enables flexible quantization while preserving information integrity.

→ Learnable Equivalent Transformation (LET) optimizes quantization boundaries for enhanced performance.

→ Distributed Quantization Calibration (DQC) strategy stabilizes training of quantized parameters.

-----

💡 Key Insights:

→ Model simplification can maintain performance while reducing complexity

→ Adaptive quantization boundaries are crucial for preserving model quality

→ Distributed calibration improves training stability

-----

📊 Results:

→ 8-bit and 6-bit models achieve comparable visual results to full-precision models

→ Significant advantages over recent leading low-bit quantization methods