Q Diffusion
Q-Diffusion is a training-free post-training quantization method designed specifically for diffusion models to accelerate image synthesis while reducing memory and computational costs. Presented at ICCV 2023, it addresses unique challenges in diffusion models such as changing output distributions across time steps and bimodal activation distributions in shortcut layers. The approach utilizes timestep-aware calibration and split shortcut quantization to compress full-precision noise estimation networks into 4-bit weights without significant performance degradation, maintaining competitive Fréchet Inception Distance scores compared to unquantized models. It supports both unconditional diffusion models like Latent Diffusion Models on datasets such as CIFAR-10 and LSUN, as well as text-guided generation frameworks like Stable Diffusion, where it achieves high-quality image generation at 4-bit weight precision. The implementation is compatible with NVIDIA TensorRT and provides calibrated quantized checkpoints for