Post-training Quantization for Hybrid Iterative Generative Models

2026-08-14

AI Summary

This study addresses the critical issues of training instability and quality degradation encountered during post-training quantization of hybrid iterative generation models. To overcome these challenges, we propose HyGenQ, a novel framework that innovatively employs hierarchical clustering to decouple outliers and integrates a scaling factor recalibration mechanism. This approach effectively suppresses error amplification while preventing truncation distortion. Extensive experiments demonstrate that HyGenQ significantly outperforms existing baselines under W8A8 post-training quantization settings and exhibits robust generalizability across diverse model families. Consequently, this work successfully enables efficient 8-bit quantized acceleration for hybrid iterative generation models without compromising generative performance, offering a practical solution for deploying large-scale generative systems with reduced computational overhead.

Abstract

Iterative Generative Models (IGMs) span autoregressive and diffusion paradigms, and hybrid variants that couple them can achieve remarkable image-generation fidelity. However, their iterative inference incurs substantial computational overhead, making Post-training Quantization (PTQ) appealing for acceleration, while directly applying vanilla PTQ to hybrid IGMs can trigger model collapse. By analyzing these failures, we identify two critical challenges: Excessive Outliers (EOs) in the activations create an irreconcilable trade-off between preserving normal precision and covering EOs, resulting in severe degradation in generation quality; Amplified Anomalies (AAs) arising unpredictably from minor quantization errors, create a mismatch between calibration and inference, thus iteratively triggering model collapse. To address these challenges, we introduce HyGenQ, a PTQ framework for hybrid IGMs. HyGenQ comprises Hierarchical Cluster Decoupling (HCD) and Scaling Recalibration (SR). HCD identifies and decouples outlier channels via a multi-stage clustering process, effectively isolating EOs while maintaining normal value precision, thereby alleviating performance degradation. SR scales AAs beyond Gaussian Bound, thereby avoiding model collapse caused by aggressive truncation. Extensive experiments demonstrate that HyGenQ successfully quantizes representative hybrid IGMs to 8-bit precision (W8A8), significantly outperforming existing baselines and validating its robustness across different model families.

Problem

• Post-training Quantization
• Hybrid Iterative Generative Models
• Model Collapse
• Excessive Outliers
• Amplified Anomalies

Innovation

• Post-training Quantization
• Hybrid Iterative Generative Models
• Hierarchical Cluster Decoupling
• Scaling Recalibration
• Model Collapse