Q-Drift: Quantization-Aware Drift Correction for Diffusion Model Sampling

TL;DR

Q-Drift is a novel sampler correction technique that mitigates quantization noise in diffusion models, significantly improving image generation quality with minimal computational overhead.

Contribution

It introduces a principled, plug-and-play drift correction method for quantized diffusion models, requiring minimal calibration data and compatible with various samplers and PTQ techniques.

Findings

Q-Drift improves FID scores across multiple models and samplers.

It requires only 5 calibration runs for effective correction.

Q-Drift preserves CLIP scores while enhancing image quality.

Abstract

Post-training quantization (PTQ) is a practical path to deploy large diffusion models, but quantization noise can accumulate over the denoising trajectory and degrade generation quality. We propose Q-Drift, a principled sampler-side correction that treats quantization error as an implicit stochastic perturbation on each denoising step and derives a marginal-distribution-preserving drift adjustment. Q-Drift estimates a timestep-wise variance statistic from calibration, in practice requiring as few as 5 paired full-precision/quantized calibration runs. The resulting sampler correction is plug-and-play with common samplers, diffusion models, and PTQ methods, while incurring negligible overhead at inference. Across six diverse text-to-image models (spanning DiT and U-Net), three samplers (Euler, flow-matching, DPM-Solver++), and two PTQ methods (SVDQuant, MixDQ), Q-Drift improves FID over the corresponding quantized baseline in most settings, with up to 4.59 FID reduction on PixArt-Sigma (SVDQuant W3A4), while preserving CLIP scores.