A training-free method for compressing diffusion models to low-bit precision (8-bit or 4-bit) by calibrating quantization parameters per denoising timestep. Because a diffusion model's activation distributions shift across the reverse denoising trajectory, applying a single set of quantization ranges degrades quality; time-aware calibration instead selects calibration data and clipping ranges tailored to each timestep, keeping full-precision-level fidelity without any retraining and extending post-training quantization to high-resolution text-guided image generation.
Proposes PQD, a training-free post-training quantization framework for diffusion models. Because a diffusion model's activation statistics vary across denoising timesteps, PQD performs time-aware cal…