RobuQ: Pushing DiTs to W1.58A2 via Robust Activation Quantization

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Yang, Kaicheng, Zhang, Xun, Qin, Haotong, Lin, Yucheng, Yang, Kaisen, Yan, Xianglong, Zhang, Yulun
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914586127499264
author Yang, Kaicheng
Zhang, Xun
Qin, Haotong
Lin, Yucheng
Yang, Kaisen
Yan, Xianglong
Zhang, Yulun
author_facet Yang, Kaicheng
Zhang, Xun
Qin, Haotong
Lin, Yucheng
Yang, Kaisen
Yan, Xianglong
Zhang, Yulun
contents Diffusion Transformers (DiTs) have recently emerged as a powerful backbone for image generation, demonstrating superior scalability and performance over U-Net architectures. However, their practical deployment is hindered by substantial computational and memory costs. While Quantization-Aware Training (QAT) has shown promise for U-Nets, its application to DiTs faces unique challenges, primarily due to the sensitivity and distributional complexity of activations. In this work, we identify activation quantization as the primary bottleneck for pushing DiTs to extremely low-bit settings. To address this, we propose a systematic QAT framework for DiTs, named RobuQ. We start by establishing a strong ternary weight (W1.58A4) DiT baseline. Building upon this, we propose RobustQuantizer to achieve robust activation quantization. Our theoretical analyses show that the Hadamard transform can convert unknown per-token distributions into per-token normal distributions, providing a strong foundation for this method. Furthermore, we propose AMPN, the first Activation-only Mixed-Precision Network pipeline for DiTs. This method applies ternary weights across the entire network while allocating different activation precisions to each layer to eliminate information bottlenecks. Through extensive experiments on unconditional and conditional image generation, our RobuQ framework achieves state-of-the-art performance for DiT quantization in sub-4-bit quantization configuration. To the best of our knowledge, RobuQ is the first achieving stable and competitive image generation on large datasets like ImageNet-1K with activations quantized to average 2 bits. The code and models will be available at https://github.com/racoonykc/RobuQ .
format Preprint
id arxiv_https___arxiv_org_abs_2509_23582
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle RobuQ: Pushing DiTs to W1.58A2 via Robust Activation Quantization
Yang, Kaicheng
Zhang, Xun
Qin, Haotong
Lin, Yucheng
Yang, Kaisen
Yan, Xianglong
Zhang, Yulun
Computer Vision and Pattern Recognition
Diffusion Transformers (DiTs) have recently emerged as a powerful backbone for image generation, demonstrating superior scalability and performance over U-Net architectures. However, their practical deployment is hindered by substantial computational and memory costs. While Quantization-Aware Training (QAT) has shown promise for U-Nets, its application to DiTs faces unique challenges, primarily due to the sensitivity and distributional complexity of activations. In this work, we identify activation quantization as the primary bottleneck for pushing DiTs to extremely low-bit settings. To address this, we propose a systematic QAT framework for DiTs, named RobuQ. We start by establishing a strong ternary weight (W1.58A4) DiT baseline. Building upon this, we propose RobustQuantizer to achieve robust activation quantization. Our theoretical analyses show that the Hadamard transform can convert unknown per-token distributions into per-token normal distributions, providing a strong foundation for this method. Furthermore, we propose AMPN, the first Activation-only Mixed-Precision Network pipeline for DiTs. This method applies ternary weights across the entire network while allocating different activation precisions to each layer to eliminate information bottlenecks. Through extensive experiments on unconditional and conditional image generation, our RobuQ framework achieves state-of-the-art performance for DiT quantization in sub-4-bit quantization configuration. To the best of our knowledge, RobuQ is the first achieving stable and competitive image generation on large datasets like ImageNet-1K with activations quantized to average 2 bits. The code and models will be available at https://github.com/racoonykc/RobuQ .
title RobuQ: Pushing DiTs to W1.58A2 via Robust Activation Quantization
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2509.23582