Stability AI 发布 Stable Diffusion 3

2024 年 2 月，Stability AI 公布了 Stable Diffusion 3。自 2022 年 Stable Diffusion 定义了开源图像生成赛道以来，这是最重要的一次架构换代。SD3 采用新的多模态扩散变压器架构，在文字渲染和复杂提示词遵循上相比前代有肉眼可见的进步——此前 AI 生成图里的文字总是歪歪扭扭，SD3 把这个问题基本解决了。

SD3 面临的竞争环境已经和两年前完全不同。闭源阵营里，Midjourney 持续迭代、DALL·E 3 与 GPT-4 深度绑定；开源阵营里，黑森林实验室的 Flux 等新对手虎视眈眈。Stability 需要一场架构上的自我革命来证明自己依然是开源图像生成的技术领导者，SD3 就是这场革命的答卷。

新架构带来的提升是真实的。多主体遵循——让模型同时正确生成多个不同对象，以及文字渲染的准确性，都是图像生成领域最难的痛点。SD3 在这两点的进步，让开源社区看到了追赶甚至超越闭源模型的可能性。开发者社区迅速围绕 SD3 展开适配，各种基于它的工作流与微调模型陆续出现。

不过，SD3 的发布也伴随着现实的阴影。Stability 在开源策略上出现了摇摆——部分权重延迟开放、许可证条款引发争议，公司自身的财务状况也一直承压。这让社区对 SD3 的感情变得复杂：既为技术进步兴奋，又为开源精神的松动担忧。

回看 SD3，它技术上的价值是清晰的——它把开源图像生成从 U-Net 时代推进到了扩散变压器时代，为后续所有新一代图像模型提供了参照。而它在商业与开源之间的拉扯，同样真实地记录了一家开源明星公司在大模型商业化浪潮中的挣扎。技术与现实两条线交织，构成了 Stable Diffusion 3 这个复杂却重要的注脚。

In February 2024 Stability AI unveiled Stable Diffusion 3. Since Stable Diffusion defined open image generation in 2022, this was the most important architecture upgrade. SD3 used a new multimodal diffusion-transformer architecture, delivering visible progress in text rendering and complex prompt adherence over the previous generation—the text in AI-generated images, once reliably crooked, was basically fixed.

SD3's competitive environment was completely different from two years earlier. In the closed camp, Midjourney kept iterating and DALL·E 3 was deeply bound to GPT-4; in the open camp, new rivals like Black Forest Labs' Flux loomed. Stability needed a self-revolution in architecture to prove it remained the technical leader of open image generation—SD3 was the answer to that test.

The gains from the new architecture were real. Multi-subject adherence—correctly generating several distinct objects at once—and text-rendering accuracy are among the hardest pain points in image generation. SD3's progress on both gave the open community hope of catching up or even surpassing closed models. Developers quickly adapted around SD3, and workflows and finetunes proliferated.

Yet SD3's release came with a shadow of reality. Stability wavered on open-source policy—some weights were delayed, license terms drew controversy, and the company's finances stayed under pressure. This left the community with mixed feelings: excitement at technical progress, worry at the loosening of the open-source spirit.

Looking back, SD3's technical value is clear—it moved open image generation from the U-Net era into the diffusion-transformer era, providing a reference for every next-generation image model. Its tug-of-war between commerce and openness equally records, truthfully, an open-source star company's struggle amid the large-model commercialization wave. Technology and reality, two threads intertwined, form the complex yet important footnote of Stable Diffusion 3.