Black Forest Labs 发布 FLUX.1

2024 年 8 月 1 日，Black Forest Labs 在宣布公司成立的同一天发布 FLUX.1。团队没有只给出一个模型名，而是摆出三扇门：效果最完整的 Pro 通过 API 提供；由 Pro 蒸馏而来的 Dev 开放权重，但面向非商业用途；追求少步快速生成的 Schnell 同样可下载，并使用 Apache 2.0 许可证。能力、速度和使用权从首日就被拆开。

这三档共享约 120 亿参数的基础架构。官方说明将多模态与并行 diffusion transformer 模块结合，并以 flow matching 作为训练基础；旋转位置编码等设计服务于效率与尺度扩展。对普通创作者，这些名词最终会被压缩成更实际的问题：复杂提示里的多个对象是否都出现，画面中的文字是否少一些乱码，同一批生成里有多少张可以直接继续编辑。

Black Forest Labs 的创始团队曾参与 VQGAN、潜空间扩散和 Stable Diffusion 等工作，这使 FLUX.1 在发布前就带着技术谱系。SDXL 长期是本地高质量文生图的默认选择，社区也因此能迅速判断新模型是否值得改工作流。FLUX.1 不需要先教育用户什么是检查点；它只需要让现有的 ComfyUI 节点、托管平台与推理脚本开始支持新的文件。

发布方称 Pro 与 Dev 在视觉质量、提示遵循、排版和输出多样性等方面超过多个商业模型，并称 Schnell 在少步模型中领先。这些是 Black Forest Labs 自己的比较，任务、评审与版本都由发布方定义。社区随后大量试图复现“更听提示”的体验，但流行速度并不等于每一种风格、语言和硬件上的质量都已得到科学共识。

许可证比排行榜更直接地决定采用路径。个人可以用 Schnell 做本地开发，研究者可以检查与修改 Dev 权重，商业产品若要使用更完整的 Pro 能力则需走 API。把三者笼统叫作“开源 Flux”会抹去最重要的边界：权重可下载、允许研究、允许商用和代码开源并不是同一件事。FLUX.1 把这些差异做成了产品线本身。

发布后的迁移发生得很快，因为它不要求社区放弃原有工具，只要求换一个模型入口。默认检查点的改变看似只是文件夹里多出几个名字，背后却是一种新的分发秩序：最强能力留在服务端，接近它的权重开放研究，快速版本用宽松许可扩散。FLUX.1 抬高的不只是图像质量，也包括开发者阅读许可证的必要性。

Black Forest Labs announced its company and released FLUX.1 on 1 August 2024. It did not offer a single model under one set of terms. The launch opened three doors. Pro, the highest-capability version, was available through APIs. Dev, distilled from Pro, shipped as downloadable weights for non-commercial use. Schnell, optimized for fast few-step generation, was also downloadable and used the Apache 2.0 license. Capability, speed, and permission were separated from the first day.

The variants shared a foundation of roughly twelve billion parameters. Black Forest Labs described a hybrid of multimodal and parallel diffusion-transformer blocks trained with flow matching, with choices such as rotary positional embeddings intended to support efficiency and scale. For a working creator, those terms eventually condensed into practical tests: whether a complex prompt preserved all its objects, whether lettering inside the image contained fewer errors, and how many images from a batch were good enough to continue editing.

The founding team carried a recognizable technical lineage, including work on VQGAN, latent diffusion, and Stable Diffusion. That history gave FLUX.1 an audience able to test it immediately. SDXL had remained a default for high-quality local text-to-image workflows, so the community did not need to learn what a checkpoint was. It needed ComfyUI nodes, hosted services, and inference scripts to accept a new set of files.

Black Forest Labs reported that Pro and Dev surpassed several commercial systems in visual quality, prompt following, typography, and output diversity, and described Schnell as a leading few-step model. These were the publisher’s comparisons, with tasks, raters, and versions chosen by the company. The community rapidly tested the claim that FLUX followed instructions more closely, but fast adoption was not scientific consensus across every style, language, and hardware configuration.

Licensing determined the adoption path more directly than the leaderboard did. An individual could use Schnell for local development. Researchers could inspect and modify Dev weights under non-commercial terms. A commercial product seeking the full Pro capability had to use an API. Calling all three “open-source Flux” erased the release’s most important boundary. Downloadable weights, permission to research, permission to commercialize, and open-source code were not interchangeable properties. FLUX.1 made their differences part of the product family.

Migration happened quickly because the release did not require the community to abandon its existing tools; users could replace the model entry inside familiar workflows. A few new filenames in a folder looked like a minor change. Behind them was a new distribution order: the strongest model stayed behind a service, a nearby distilled model opened for research, and a fast variant spread under a permissive license. FLUX.1 raised the image-quality ceiling and the cost of failing to read the license.