OpenAI 发布 DALL·E

2021 年 1 月 5 日，OpenAI 的博客标题很老实：DALL·E: Creating images from text。没有可随意注册的产品入口，读者看到的是一组挑选过的生成图，以及一条技术路线说明。模型大约 120 亿参数；名字把达利的幻觉与 WALL·E 的机械拼在一起——营销可以玩梗，正文仍要把机制写清。

机制并不神秘，只是当时少见。图像先被离散变分自编码器压成 token，文本也是 token；两者排进同一条最长约 1280 的序列（大约 256 个文本位置与 1024 个图像位置），由解码器 Transformer 自回归预测。训练目标仍是「下一个 token」，只不过后半段 token 对应的是图像块而不是单词。读者若刚读完 GPT-3，会有熟悉的刺痛：规模化的语言建模，开始直接在视觉离散符号上续写。

博客样例比公式更有传播力。鳄梨形状的扶手椅、文字指令下的几何变形、把无关概念硬捏在同一物体上的组合——这些图证明模型不只是在检索训练集里最像的照片，而能在表示空间里做某种拼接。必须同时写下边界：公开页面展示的是团队愿意放出的成功与趣味案例，不是任意提示的平均分布，也不是可商用 API 的服务等级协议。DALL·E 当时首先是研究演示与议程设置，不是 Photoshop 的替代品上架。

同日前后，OpenAI 还把 CLIP 推到台前：用图文对学习可对齐的嵌入空间，擅长检索与零样本分类，而不是直接画像素。两条线后来会缠在一起——许多文生图系统用类 CLIP 文本塔当条件——但在 2021 年 1 月，它们是分工不同的实验：一个问「语言如何生成图像 token」，一个问「语言与图像如何放进同一度量」。把两者混成「OpenAI 当天发明了文生图产品」，会抹掉演示与工程化之间的距离。CLIP 像尺子，DALL·E 像画笔；后来的系统常常一只手握尺子、一只手握画笔，但首日材料里它们是分开的章节。

对行业议程的影响，写在随后两年的对照里。实验室开始把「提示控制图像」当作必须公开回应的能力；DALL·E 2、Imagen、Parti 等以更高分辨率与更稳的构图跟进；2022 年的 Stable Diffusion 则把可下载权重丢进消费级显卡，把演示特权拆成文件分发。版权、训练数据与深度伪造的争论几乎与样例同步出现——能力一旦可被一句话唤起，治理问题就不肯等产品成熟再排队。

离散 token 路线也留下工程遗产与局限。把连续像素切成码本索引，让语言模型工具链可以直接接手，却也把纹理细节与文字清晰度卡在码本精度上；后续扩散模型在连续或潜空间里迭代去噪，部分是为了绕开这条瓶颈。回到那篇博客本身：它没有承诺人人都能立刻画自己的品牌吉祥物，它展示的是——在足够大的自回归模型里，文本条件可以成为图像序列的前缀。后来的扩散路线会改掉采样方式，产品会改掉访问方式，但「用自然语言点名画面」这一用户动作，在 2021 年初已经被写成可截图传播的事实。

On 5 January 2021 OpenAI’s blog title stayed literal: DALL·E: Creating images from text. There was no open signup queue. Readers got curated generations and a short technical map. The model sat near 12 billion parameters; the name joked Dalí with WALL·E. Marketing could pun. The mechanism still had to be written cleanly.

It was not mysterious—only uncommon then. A discrete VAE turned images into tokens; text was already tokens. Both entered one sequence on the order of 1,280 positions (roughly 256 text and 1,024 image), predicted by a decoder-only Transformer. The training objective remained next-token prediction; the later tokens simply stood for image patches instead of words. Anyone fresh from GPT-3 felt a familiar sting: scaled language modeling had begun to continue visual discrete symbols.

The blog samples traveled farther than the equations. An avocado armchair, geometric edits under verbal instruction, unrelated concepts forced into one object—evidence that the system was not only retrieving nearest training photos, but composing in representation space. Boundaries belong beside the spectacle: the page showed what the team chose to publish, not the average of arbitrary prompts, not an SLA for a commercial API. In early 2021 DALL·E was research demo and agenda-setting, not a Photoshop substitute on the shelf.

Around the same day OpenAI also put CLIP forward: image–text pairs learning an aligned embedding space for retrieval and zero-shot classification, not direct pixel painting. The two lines would later entangle—many text-to-image stacks condition on CLIP-like text towers—but in January they were different experiments: one asked how language can generate image tokens; the other asked how language and images share a metric. Collapsing both into “OpenAI shipped a text-to-image product that day” erases the distance between demo and productization. CLIP is a ruler; DALL·E is a brush. Later systems often hold both; the first-day materials keep them in separate chapters.

What changed was the public checklist for labs. Prompt-controlled imagery became something you had to answer in public. DALL·E 2, Imagen, Parti and peers chased resolution and composition; Stable Diffusion in 2022 dropped downloadable weights onto consumer GPUs and turned demo privilege into file distribution. Copyright, training data, and deepfake arguments arrived almost as soon as the samples did—once a sentence could summon a picture, governance refused to wait for mature products.

The discrete-token route left engineering inheritance and limits. Codebook indices let language-model tooling take over, but texture and text clarity hit codebook precision; later diffusion systems partly exist to step around that bottleneck. Return to that first blog. It did not promise every brand would mint mascots overnight. It showed that, inside a large enough autoregressive model, text can prefix an image token stream. Later diffusion systems would change sampling; products would change access. The user gesture—naming a picture in natural language—was already a screenshotable fact in early 2021.