Stable Diffusion 开放文生图权重

大约 4GB 的权重文件，落在一块当时常见的消费级显卡上，就能本地出图。

2022 年 8 月前后，Stability AI 与 CompVis 等合作方放出 Stable Diffusion 的代码与权重（公众广泛使用的 1.4 检查点约这一量级）。模型不在像素空间里硬扛完整扩散，而在压缩后的潜空间去噪：自动编码器先把图像压到低维潜变量，U-Net 在那里迭代，再解码回 512×512 像素。文本条件通过预训练的 CLIP 文本编码器注入。对用户最直接的后果是：推理栈第一次像「能装进游戏玩家主机的软件」。

对照很清楚。OpenAI 的 DALL·E 2、Google 的 Imagen 已经展示了高质量文本引导图像，但主要经由封闭服务或受控演示提供，开发者难以检查权重、改采样器或在私有数据上继续训练。Stable Diffusion 把推理摊开：可以换采样步数、换调度器、接负向提示，也可以把权重当初始化做领域微调。许可与训练数据版权问题随即成为长期争议——「可下载」不等于「训练语料已获全面授权」，也不等于 OSI 意义上的无限制开源；更准确的说法是开放权重与可运行代码释放了改造权。

社区速度超过了任何单一实验室的发布节奏。Automatic1111 等 Web UI 降低了使用门槛；ControlNet 把边缘、深度、姿态等结构条件接进生成；LoRA 与其它低秩适配让人用很少显存就练出角色或画风；模型合并、画廊网站与提示词分享形成平行生态。与此同时，非自愿色情图像、艺术家风格仿冒、深度伪造与平台治理问题同步爆发。开放没有自动附赠安全方案；它只是把能力的分发曲线拉平。

训练数据来自大规模图文对（如 LAION 相关集合），过滤与安全策略在发布说明与后续版本中多次调整，本身即是争议焦点。推理侧常用 20–50 步采样、分类器自由引导（CFG）调节提示遵循度——这些旋钮进入社区教程后，用户不必理解完整扩散推导，也能稳定出图。潜空间压缩比决定细节与速度的权衡：压缩越狠，显存越省，细纹理与小文字越容易糊；社区随后用更高分辨率检查点与专用适配器修补这些弱点。

之后的 SDXL、视频模型与闭源旗舰会继续改写质量上限，但 2022 年这件事的独特性在于分发：权重一旦落到个人电脑，微调与界面创新就不再需要向模型提供方申请额度。Stable Diffusion 没有结束关于版权与同意的争论；它让争论有了具体的文件哈希与本地路径。生成图像的历史，从此必须同时写实验室论文与社区仓库提交记录。

Roughly 4GB of weights, on a then-common consumer GPU, was enough to generate images locally.

Around August 2022 Stability AI, CompVis, and collaborators released Stable Diffusion code and weights (the widely used 1.4 checkpoint on that order). The model does not run full diffusion in pixel space. It denoises in a compressed latent space: an autoencoder maps images to low-dimensional latents, a U-Net iterates there, and a decoder returns 512×512 pixels. Text conditioning enters through a pretrained CLIP text encoder. The immediate user consequence: the inference stack first felt like software that fits a gamer PC.

The contrast is clear. OpenAI’s DALL·E 2 and Google’s Imagen already showed strong text-guided imagery, mainly through closed services or controlled demos. Developers could not inspect weights, swap samplers, or continue training on private data easily. Stable Diffusion opened the stack: change steps and schedulers, attach negative prompts, treat weights as initialization for domain fine-tunes. License terms and training-data copyright became lasting disputes—“downloadable” is not “training corpus fully cleared,” nor OSI open source without restriction. The accurate claim is that open weights and runnable code released the right to modify.

Community speed outran any single lab’s release cadence. Web UIs such as Automatic1111 lowered the floor; ControlNet wired edges, depth, pose, and other structure into generation; LoRA and related low-rank adapters let people train characters or styles on little VRAM; model merging, gallery sites, and prompt sharing formed a parallel ecosystem. At the same time non-consensual sexual imagery, artist-style mimicry, deepfakes, and platform governance problems erupted. Openness does not ship a safety plan; it flattens the distribution curve of capability.

Training data from large image–text collections, including LAION-related sets, made filtering and safety policy permanent flashpoints. Inference knobs—step counts around 20–50, classifier-free guidance—entered community tutorials so users could ship images without deriving the full math. Latent compression trades detail for memory; later checkpoints and adapters patched texture and small-text rendering.

Later SDXL, video models, and closed flagships rewrite quality ceilings; 2022’s distinctive fact is distribution. Once weights land on personal machines, fine-tunes and UI invention no longer need quota from the model provider. Stable Diffusion did not end arguments about copyright and consent; it gave those arguments concrete file hashes and local paths. The history of generative images must now be written from lab papers and community repository commits together.