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

2026-03-18

AI Summary

This work addresses the degradation in generation quality caused by the accumulation of quantization noise during the denoising process of diffusion models. To mitigate this issue, the authors propose Q-Drift, a novel method that, for the first time, models quantization error as an implicit stochastic perturbation along the diffusion trajectory and derives a drift correction term that preserves the marginal distribution. Q-Drift is plug-and-play, requiring no retraining; it only needs a small set of calibration samples to estimate per-timestep variance statistics. It seamlessly integrates with mainstream samplers (e.g., Euler, flow-matching, DPM-Solver++), model architectures (DiT, U-Net), and post-training quantization schemes (SVDQuant, MixDQ). Evaluated across six text-to-image models, Q-Drift significantly improves FID—by up to 4.59 on PixArt-Sigma with SVDQuant W3A4—while maintaining CLIP scores.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Calibration & Uncertainty Quantification
• Knowledge Representation and Reasoning: Qualitative Reasoning

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

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.

Problem

• post-training quantization
• quantization noise
• diffusion model
• generation quality
• denoising trajectory

Innovation

• quantization-aware correction
• diffusion model sampling
• post-training quantization
• drift adjustment
• marginal-distribution preservation