DiG: Scalable and Efficient Diffusion Models with Gated Linear Attention

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Zhu, Lianghui, Huang, Zilong, Liao, Bencheng, Liew, Jun Hao, Yan, Hanshu, Feng, Jiashi, Wang, Xinggang
Natura: Preprint
Pubblicazione: 2024
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866912134152060928
author Zhu, Lianghui
Huang, Zilong
Liao, Bencheng
Liew, Jun Hao
Yan, Hanshu
Feng, Jiashi
Wang, Xinggang
author_facet Zhu, Lianghui
Huang, Zilong
Liao, Bencheng
Liew, Jun Hao
Yan, Hanshu
Feng, Jiashi
Wang, Xinggang
contents Diffusion models with large-scale pre-training have achieved significant success in the field of visual content generation, particularly exemplified by Diffusion Transformers (DiT). However, DiT models have faced challenges with quadratic complexity efficiency, especially when handling long sequences. In this paper, we aim to incorporate the sub-quadratic modeling capability of Gated Linear Attention (GLA) into the 2D diffusion backbone. Specifically, we introduce Diffusion Gated Linear Attention Transformers (DiG), a simple, adoptable solution with minimal parameter overhead. We offer two variants, i,e, a plain and U-shape architecture, showing superior efficiency and competitive effectiveness. In addition to superior performance to DiT and other sub-quadratic-time diffusion models at $256 \times 256$ resolution, DiG demonstrates greater efficiency than these methods starting from a $512$ resolution. Specifically, DiG-S/2 is $2.5\times$ faster and saves $75.7\%$ GPU memory compared to DiT-S/2 at a $1792$ resolution. Additionally, DiG-XL/2 is $4.2\times$ faster than the Mamba-based model at a $1024$ resolution and $1.8\times$ faster than DiT with FlashAttention-2 at a $2048$ resolution. We will release the code soon. Code is released at https://github.com/hustvl/DiG.
format Preprint
id arxiv_https___arxiv_org_abs_2405_18428
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle DiG: Scalable and Efficient Diffusion Models with Gated Linear Attention
Zhu, Lianghui
Huang, Zilong
Liao, Bencheng
Liew, Jun Hao
Yan, Hanshu
Feng, Jiashi
Wang, Xinggang
Computer Vision and Pattern Recognition
Artificial Intelligence
Diffusion models with large-scale pre-training have achieved significant success in the field of visual content generation, particularly exemplified by Diffusion Transformers (DiT). However, DiT models have faced challenges with quadratic complexity efficiency, especially when handling long sequences. In this paper, we aim to incorporate the sub-quadratic modeling capability of Gated Linear Attention (GLA) into the 2D diffusion backbone. Specifically, we introduce Diffusion Gated Linear Attention Transformers (DiG), a simple, adoptable solution with minimal parameter overhead. We offer two variants, i,e, a plain and U-shape architecture, showing superior efficiency and competitive effectiveness. In addition to superior performance to DiT and other sub-quadratic-time diffusion models at $256 \times 256$ resolution, DiG demonstrates greater efficiency than these methods starting from a $512$ resolution. Specifically, DiG-S/2 is $2.5\times$ faster and saves $75.7\%$ GPU memory compared to DiT-S/2 at a $1792$ resolution. Additionally, DiG-XL/2 is $4.2\times$ faster than the Mamba-based model at a $1024$ resolution and $1.8\times$ faster than DiT with FlashAttention-2 at a $2048$ resolution. We will release the code soon. Code is released at https://github.com/hustvl/DiG.
title DiG: Scalable and Efficient Diffusion Models with Gated Linear Attention
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2405.18428