SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Xie, Enze, Chen, Junsong, Chen, Junyu, Cai, Han, Tang, Haotian, Lin, Yujun, Zhang, Zhekai, Li, Muyang, Zhu, Ligeng, Lu, Yao, Han, Song
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909356375670784
author Xie, Enze
Chen, Junsong
Chen, Junyu
Cai, Han
Tang, Haotian
Lin, Yujun
Zhang, Zhekai
Li, Muyang
Zhu, Ligeng
Lu, Yao
Han, Song
author_facet Xie, Enze
Chen, Junsong
Chen, Junyu
Cai, Han
Tang, Haotian
Lin, Yujun
Zhang, Zhekai
Li, Muyang
Zhu, Ligeng
Lu, Yao
Han, Song
contents We introduce Sana, a text-to-image framework that can efficiently generate images up to 4096$\times$4096 resolution. Sana can synthesize high-resolution, high-quality images with strong text-image alignment at a remarkably fast speed, deployable on laptop GPU. Core designs include: (1) Deep compression autoencoder: unlike traditional AEs, which compress images only 8$\times$, we trained an AE that can compress images 32$\times$, effectively reducing the number of latent tokens. (2) Linear DiT: we replace all vanilla attention in DiT with linear attention, which is more efficient at high resolutions without sacrificing quality. (3) Decoder-only text encoder: we replaced T5 with modern decoder-only small LLM as the text encoder and designed complex human instruction with in-context learning to enhance the image-text alignment. (4) Efficient training and sampling: we propose Flow-DPM-Solver to reduce sampling steps, with efficient caption labeling and selection to accelerate convergence. As a result, Sana-0.6B is very competitive with modern giant diffusion model (e.g. Flux-12B), being 20 times smaller and 100+ times faster in measured throughput. Moreover, Sana-0.6B can be deployed on a 16GB laptop GPU, taking less than 1 second to generate a 1024$\times$1024 resolution image. Sana enables content creation at low cost. Code and model will be publicly released.
format Preprint
id arxiv_https___arxiv_org_abs_2410_10629
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers
Xie, Enze
Chen, Junsong
Chen, Junyu
Cai, Han
Tang, Haotian
Lin, Yujun
Zhang, Zhekai
Li, Muyang
Zhu, Ligeng
Lu, Yao
Han, Song
Computer Vision and Pattern Recognition
We introduce Sana, a text-to-image framework that can efficiently generate images up to 4096$\times$4096 resolution. Sana can synthesize high-resolution, high-quality images with strong text-image alignment at a remarkably fast speed, deployable on laptop GPU. Core designs include: (1) Deep compression autoencoder: unlike traditional AEs, which compress images only 8$\times$, we trained an AE that can compress images 32$\times$, effectively reducing the number of latent tokens. (2) Linear DiT: we replace all vanilla attention in DiT with linear attention, which is more efficient at high resolutions without sacrificing quality. (3) Decoder-only text encoder: we replaced T5 with modern decoder-only small LLM as the text encoder and designed complex human instruction with in-context learning to enhance the image-text alignment. (4) Efficient training and sampling: we propose Flow-DPM-Solver to reduce sampling steps, with efficient caption labeling and selection to accelerate convergence. As a result, Sana-0.6B is very competitive with modern giant diffusion model (e.g. Flux-12B), being 20 times smaller and 100+ times faster in measured throughput. Moreover, Sana-0.6B can be deployed on a 16GB laptop GPU, taking less than 1 second to generate a 1024$\times$1024 resolution image. Sana enables content creation at low cost. Code and model will be publicly released.
title SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2410.10629