按年被提到
20212022 · 3 次2026
视觉、语音与图像生成›扩散模型·arXiv:2112.10752 · cs.CV

High-Resolution Image Synthesis with Latent Diffusion Models

基于潜在扩散模型的高分辨率图像合成
Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, Björn Ommer · 2021

在自编码器潜空间里做扩散,Stable Diffusion 的原论文

开始阅读 ↗PDFalphaXiv 讨论
被推荐 · 8 人首次 2022-08-17 · 最近 2025-12-30发表于 IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings
8不同的人
8活跃月份
9讨论串
16,800被引(OpenAlex)

论文不在像素空间而在预训练自编码器的潜空间中训练扩散模型,在复杂度降低与细节保留之间取得良好折中,大幅减少训练和推理的算力需求。通过引入交叉注意力层,模型可接受文本、边界框等通用条件输入,并能以卷积方式做高分辨率合成。潜在扩散模型在图像修复上达到当时最优,在无条件生成、语义场景合成和超分辨率上也具有竞争力。

HN 为什么提它

HN 上的人把它当作 Stable Diffusion 背后的原始论文:说明“在潜空间做扩散”如何大幅降低算力,从而让图像生成走向大众。

奠基之作难度 4 / 5约 60 分钟经得起时间
读之前最好知道
1扩散模型(diffusion models)
2变分自编码器(VAE)
3U-Net 架构
4交叉注意力(cross-attention)
5生成模型基础

HN 书评

推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及
推荐godelski · 2023-07

“you need these tools to understand Stable Diffusion which really is just Latent Diffusion[4]”

指出理解 Stable Diffusion 需要这些工具,它本质就是 Latent Diffusion
Stanford A.I. Courses原评论 ↗
批评godelski · 2024-07

“Many of you may have experienced this when the first stable diffusion paper came out and the images in the paper were incredible but when you used the hugging face generator you'd get nothing nearly as good.”

AI models collapse when trained on recursively generated data原评论 ↗
引用lucidrains · 2022-08

“They continued building off their latent diffusion direction (encode with vqgan-vae and then diffusion in latent space)”

说明其路线:先用 vqgan-vae 编码,再在潜空间做 diffusion
Build Your Own Imagen Text-to-Image Model原评论 ↗
引用Dzugaru · 2022-09

“after that Latent Diffusion Models [4] democratized the thing by vastly reducing the amount of computation needed - operating in latent space”

指出 Latent Diffusion 在潜空间运算,大幅降低计算量,让技术普及
Stable Diffusion forming images from text: image snapshots at each step原评论 ↗
引用rassibassi · 2022-09

“another reason for the small model size and the novelty of the underlying paper [1], is that the diffusion model is not acting on the pixel space but rather on a latent space.”

解释模型小的原因:diffusion 作用在潜空间而非像素空间
A Web UI for Stable Diffusion原评论 ↗
引用firechickenbird · 2023-03

“Very interesting, although I was expecting the Stable Diffusion paper[1] to be one of the most cited in 2022”

原本以为 Stable Diffusion 这篇会是 2022 年被引最多的论文之一
The Most Cited AI Papers in 2022原评论 ↗
看全部 9 条 →
引用capableweb · 2023-04

“Links to High-Resolution Image Synthesis with Latent Diffusion Models which is the Version 2 of the paper, indeed released in April 2022.”

指出链接的是该论文 2022 年 4 月发布的第 2 版,Stable Diffusion 代码即源自它
Leaked deck raises questions over Stability AI’s Series A pitch to investors原评论 ↗
引用parlancex · 2024-02

“This is explained in the original latent diffusion paper on which Stable Diffusion was based: High-Resolution Image Synthesis with Latent Diffusion Models”

指向原始论文:它解释了约束 KL 损失权重是为了感知压缩而非语义压缩
The VAE Used for Stable Diffusion Is Flawed原评论 ↗
引用cookiengineer · 2025-12

“I only implemented the initial SD paper [1] back then so I have no clue whether the u-net architecture changed when it comes to the semantic mapping of the text/labels.”

提到自己当年只实现过最初的 SD 论文,不清楚 u-net 结构后来是否改变
Show HN: I remade my website in the Sith Lord Theme and I hope it's fun原评论 ↗