OpenAI 的图像生成模型家族。DALL-E 1(2021)用离散 VAE + Transformer 的方法。DALL-E 2(2022)用 CLIP + 扩散。DALL-E 3(2023)集成到 ChatGPT,强调 prompt 遵循 — 它用一个 LLM 在生成前把用户 prompt 重写成详细的图像描述,显著改善了你要求的和你得到的之间的匹配。
为什么重要
DALL-E 是让公众知道 AI 图像生成的模型。DALL-E 2 在 2022 年的发布爆红,引发了对 AI 生成图像的兴奋和担忧。DALL-E 3 与 ChatGPT 的集成让图像生成对数亿用户开放。它的 prompt 重写创新影响了其他模型如何处理 text-to-image 转换。
Deep Dive
DALL-E 3's key innovation: instead of feeding user prompts directly to the image model, it uses GPT-4 to expand vague prompts into detailed image descriptions. "A cat" becomes "A fluffy orange tabby cat sitting on a windowsill, afternoon sunlight streaming in, photorealistic style, warm tones." This prompt rewriting dramatically improves output quality because diffusion models respond better to detailed descriptions than to short prompts.
Safety Measures
DALL-E has the most aggressive safety filters in the industry: it refuses to generate images of real public figures, violent content, and sexual content. It also uses C2PA metadata (Content Credentials) to mark images as AI-generated. These safety choices limit DALL-E's flexibility compared to open alternatives (Stable Diffusion, Flux) but reflect OpenAI's approach to responsible deployment. The trade-off between safety and creative freedom is a defining tension in image generation.
API and Integration
DALL-E 3 is available through OpenAI's API and through ChatGPT. The API provides more control (image size, quality settings, style parameter) but the ChatGPT integration is more popular because it handles prompt engineering automatically. The integration model — LLM + image generator as a unified experience rather than separate tools — influenced competitors and is becoming the standard for consumer image generation.
