文生图Text-to-Image Generation

打字变图片：扩散模型最出圈的应用，从像素空间搬进潜空间才变得可用。Typing turns into pictures: diffusion's most visible application — practical only after moving from pixel space to latent space.

早期扩散模型直接在像素空间去噪，训练要烧数百 GPU 天。

latent diffusion 论文把扩散搬进预训练自编码器的潜空间。训练与推理成本大幅下降。

加入交叉注意力后，文本等条件可以直接引导生成，文生图由此成立。

该论文（Stable Diffusion 的基础）2022 年发表于 CVPR。

Early diffusion models denoised directly in pixel space, costing hundreds of GPU-days to train.

The latent diffusion paper moved diffusion into the latent space of a pretrained autoencoder, slashing training and inference cost.

Adding cross-attention let text and other conditions guide generation — making text-to-image work.

That paper, the basis of Stable Diffusion, was published at CVPR 2022.

「输入一句话出图」不是魔法，是这套架构组合。知道潜空间这步，才懂为什么它比早期方案便宜可用。'Type a sentence, get a picture' is not magic but this architecture stack. Knowing the latent-space step explains why it is affordable where early approaches were not.

去 Models 图谱：看这个模型有哪些量化版、按你的显存该下哪一档Models atlas: which quantisations exist and which tier fits your VRAM ↗

回到名词地图看第1站的其他 34 个词 →Back to the map: the other 34 terms in station 1 →

• Rombach 等《High-Resolution Image Synthesis with Latent Diffusion Models》（CVPR 2022）Rombach et al., High-Resolution Image Synthesis with Latent Diffusion Models (CVPR 2022) ↗