跳到主内容
@wquguru
精选90Hacker News Best（web_list）模型发布/更新

BFL发布5.6B参数开源图像模型FLUX 3 Image Klein

FLUX 3 Image Klein

原文
发到 X
推荐理由

FLUX 3系列全新旗舰开源发布，在显著降低参数量与硬件门槛的同时实现性能跃迁，对本地部署与开发者极具价值。

Back to blog

返回博客

  • Models
  • News
  • Research
  • 模型
  • 新闻
  • 研究

FLUX 3 Image Klein

Today we are releasing FLUX 3 Image Klein, a 5.6B parameter open weights model for text-to-image generation, image editing and multi-reference generation with up to 10 reference images. It shares core capabilities with the full FLUX 3 Image model: native 4K generation, pixel-perfect editing, and bounding boxes that control where elements are placed or confine edits to a defined area.

今天我们发布 FLUX 3 Image Klein，这是一个拥有 56 亿参数的开放权重模型，支持文本到图像生成、图像编辑以及多参考图像生成（最多支持 10 张参考图像）。它与完整的 FLUX 3 Image 模型共享核心能力：原生 4K 生成、像素级精确编辑，以及通过边界框控制元素放置位置或将编辑限制在定义区域内。

FLUX 3 Image Klein builds on FLUX 3's shared learnings across images, video and audio. Images contribute spatial detail, video adds motion and physical dynamics, and audio provides information about events and interactions. Together, these signals help the model generate images that feel grounded in the real world.

FLUX 3 Image Klein 建立在 FLUX 3 在图像、视频和音频方面共享的学习成果之上。图像提供空间细节，视频增加运动和物理动态，音频提供关于事件和交互的信息。这些信号共同帮助模型生成具有现实世界根基感的图像。

The release includes a base model for fine-tuning and a step distilled model for low-latency generation, along with FP8, INT8 and NVFP4 builds developed with NVIDIA support, under the FLUX Kommunity License.

此次发布包括用于微调的基础模型和用于低延迟生成的蒸馏模型，以及在 NVIDIA 支持下开发的 FP8、INT8 和 NVFP4 版本，均遵循 FLUX Kommunity License 许可协议。

Try it in the BFL API · Download the weights

在 BFL API 中试用 · 下载权重

Leading open weights quality with fewer parameters

以更少的参数实现领先的开放权重质量

In internal evaluations, FLUX 3 Image Klein ranks first in text-to-image, multi-reference generation and image editing, ahead of larger open weights models. It also rates above the closed models in the comparison, Ideogram V4.5 and FLUX.2 [pro]. At 5.6B parameters, it is the smallest open model in the comparison, against 7B for Qwen-Image 2.1 and 12.9B for Krea 2. Compared with FLUX.2 [klein] 9B, it gains 91 Elo in text-to-image and 70 in image editing, with 38% fewer parameters.

在内部评估中，FLUX 3 Image Klein 在文本到图像生成、多参考图像生成和图像编辑方面排名第一，优于更大的开放权重模型。在与闭源模型 Ideogram V4.5 和 FLUX.2 [pro] 的对比中，其评分也更高。凭借 56 亿参数，它是本次对比中最小的开放模型，相比之下 Qwen-Image 2.1 为 70 亿参数，Krea 2 为 129 亿参数。与 FLUX.2 [klein] 90 亿参数相比，它在文本到图像生成上提升了 91 Elo 分，在图像编辑上提升了 70 Elo 分，而参数量减少了 38%。

FLUX 3 Image Klein was also optimized for consumer GPUs. Fewer parameters make it practical to run locally and cheaper to adapt, and teams with data sovereignty or local deployment requirements can evaluate and adapt the model on their own infrastructure. This is also key for on-device image editing, allowing consumers to edit images directly on their favorite laptops. The generator weights require 11.2 GB in BF16 and 5.6 GB in FP8. With the INT8 build, FLUX 3 Image Klein runs on NVIDIA GeForce RTX GPUs with as little as 8GB of VRAM, through ComfyUI or Diffusers, extending what FLUX.2 [klein] started to even smaller hardware. With the distilled model and the INT8 build, it generates a 1MP image in about 4 seconds on an NVIDIA GeForce RTX 5090 and under 9 seconds on an RTX 4070 through ComfyUI.

FLUX 3 Image Klein 还针对消费级 GPU 进行了优化。更少的参数使其适合本地运行且适配成本更低，具有数据主权或本地部署需求的团队可以在自己的基础设施上评估和调整该模型。这对于设备端图像编辑也很关键，允许消费者直接在他们喜爱的笔记本电脑上编辑图像。生成器权重在 BF16 格式下需要 11.2 GB 显存，在 FP8 格式下需要 5.6 GB 显存。使用 INT8 版本时，FLUX 3 Image Klein 可通过 ComfyUI 或 Diffusers 在配备低至 8GB VRAM 的 NVIDIA GeForce RTX GPU 上运行，将 FLUX.2 [klein] 开创的可能性扩展至更小型的硬件。结合蒸馏模型和 INT8 版本，它可在 NVIDIA GeForce RTX 5090 上约 4 秒内生成一张 1MP 图像，在 RTX 4070 上通过 ComfyUI 用时不到 9 秒。

Generation time by GPU: Step-distilled model, 16 steps. Seconds per image, one image per run. Lower is better. Measured by NVIDIA; RTX results in ComfyUI; data results are the best-performing framework for each configuration. Each GPU shows its quickest build. Generation time on the GPU itself; API latency in Section 02 also includes queueing, network and prompt upsampling.

GPU 生成时间：步蒸馏模型，16 步。每张图像所需秒数，每次运行一张图像。数值越低越好。由 NVIDIA 测量；RTX 结果基于 ComfyUI；数据结果为每种配置下表现最佳的框架。每个 GPU 显示其最快的构建版本。此处为 GPU 本身的生成时间；第 02 节中的 API 延迟还包括排队、网络和处理提示词上采样（prompt upsampling）的时间。

Generation time by GPU: Base model, 50 steps. Seconds per image, one image per run. Lower is better. Measured by NVIDIA; RTX results in ComfyUI. Each GPU shows its quickest build.

GPU 生成时间：基础模型，50 步。每张图像所需秒数，每次运行一张图像。数值越低越好。由 NVIDIA 测量；RTX 结果基于 ComfyUI。每个 GPU 显示其最快的构建版本。

The base checkpoint is built for fine-tuning: Diffusers and Ostris's AI Toolkit support LoRA training from day one, and NVIDIA's NeMo AutoModel ships recipes that work directly from the released checkpoints.

基础检查点专为微调而构建：Diffusers 和 Ostris 的 AI Toolkit 从第一天起就支持 LoRA 训练，NVIDIA 的 NeMo AutoModel 提供了可直接使用发布检查点的配方。

For even faster production workloads that need lower latency, we also work with teams on specific customizations that get you there. Get in touch.

对于需要更低延迟的更快生产工作负载，我们还与团队一起进行特定的定制以达成目标。请联系我们要了解更多。

On the speed–quality frontier

在速度-质量前沿

FLUX 3 Image Klein also extends the speed–quality Pareto frontier. Through the BFL API, it delivers lower latency than Qwen-Image 2.1, FLUX.2 [pro], Ideogram V4.5 and Krea 2, and the highest rating at its latency and above. In the charts below, the violet line traces the highest rating available at each latency and the gray line shows the previous frontier.

FLUX 3 Image Klein 也扩展了速度-质量的帕累托前沿。通过 BFL API，它提供的延迟低于 Qwen-Image 2.1、FLUX.2 [pro]、Ideogram V4.5 和 Krea 2，并在同等及更高延迟下获得最高评分。在下图中，紫色线条追踪每个延迟级别可用的最高评分，灰色线条显示之前的前沿。

Latency is measured end to end, from request submission to the finished 1 MP image, including queueing. FLUX 3 Image Klein was timed on our servers. Other models were timed from the client, which includes polling and network time. Krea 2 was timed on its Medium tier; its Large tier took 31.1 seconds for text-to-image.

延迟是从请求提交到完成 1 MP 图像的端到端测量，包括排队时间。FLUX 3 Image Klein 在我们的服务器上计时。其他模型从客户端计时，其中包括轮询和网络时间。Krea 2 在其 Medium 层级上计时；其 Large 层级进行文生图耗时 31.1 秒。

Get started

开始使用

FLUX 3 Image Klein is available today as open weights and through the BFL API.

FLUX 3 Image Klein 现已作为开放权重提供，并通过 BFL API 提供服务。

Try it

尝试一下

  • BFL API: the distilled model is available as flux-3-image-klein, with prompt upsampling and pixel restoration, which preserves unedited regions of an image exactly. Pricing starts at $0.009 per 1K image.
  • Playground or Hugging Face.
  • BFL API：蒸馏模型名为 flux-3-image-klein，具备提示词上采样和像素还原功能，可精确保留图像未编辑的区域。价格起步价为每 1K 张图像 0.009 美元。
  • Playground 或 Hugging Face。

Build with it

构建应用

  • Weights on Hugging Face: base and distilled.
  • Reference inference code, including a Gradio demo for bounding-box control.
  • Try it out on Hugging Face Spaces.
  • Day-one support in ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni.
  • LoRA training with Ostris's AI Toolkit and recipes in NVIDIA NeMo AutoModel.
  • Trained on NVIDIA GB200 NVL72 and optimized for NVIDIA's accelerated computing platform, serving through NVIDIA TensorRT LLM, vLLM-Omni and SGLang Diffusion.
  • On NVIDIA GB300, the distilled model generates a 1MP image in under 0.6 seconds in NVFP4. Serving a LoRA adds around 5% to generation time.
  • Hugging Face 上的权重：基础模型和蒸馏模型。
  • 参考推理代码，包括用于边界框控制的 Gradio 演示。
  • 在 Hugging Face Spaces 上试用。
  • ComfyUI、Hugging Face Diffusers、TensorRT-LLM、SGLang Diffusion 和 vLLM-Omni 均提供首日支持。
  • 使用 Ostris 的 AI Toolkit 进行 LoRA 训练，并使用 NVIDIA NeMo AutoModel 中的配方。
  • 在 NVIDIA GB200 NVL72 上训练，并针对 NVIDIA 加速计算平台进行了优化，通过 NVIDIA TensorRT LLM、vLLM-Omni 和 SGLang Diffusion 提供服务。
  • 在 NVIDIA GB300 上，蒸馏模型在 NVFP4 精度下可在不到 0.6 秒内生成 1MP 图像。加载 LoRA 会使生成时间增加约 5%。

License

许可证

The weights are available under the FLUX Kommunity License, which allows free commercial use for individuals and companies with annual revenue under $5M, alongside research and non-commercial use. Larger organizations can deploy commercially through our self-hosted licensing options or API.

权重采用 FLUX Kommunity 许可证提供，允许年收入低于 500 万美元的个人和公司免费用于商业用途，同时支持研究和非商业用途。大型组织可通过我们的自托管许可选项或 API 进行商业部署。

Safety

安全性

In an effort to make AI technology both open and safe, we implement multiple safeguards throughout the model development lifecycle to combat misuse. These include dataset filters during pre-training to limit exposure to harmful content, and targeted fine-tuning during post-training to suppress harmful behaviors, making it progressively harder to elicit harmful content. Ahead of the release, the FLUX 3 Image Klein checkpoints were rigorously red teamed internally and evaluated by our external safety partner Cinder, demonstrating substantially fewer vulnerabilities than other popular open weights models. More on our safety methodology is available here.

为了使 AI 技术既开放又安全，我们在模型开发生命周期的各个环节实施了多项安全措施以防范滥用。这些措施包括在预训练阶段使用数据集过滤器以限制接触有害内容，以及在后训练阶段进行针对性微调以抑制有害行为，从而逐步增加生成有害内容的难度。在发布之前，FLUX 3 Image Klein 检查点已通过内部严格的红队测试，并由我们的外部安全合作伙伴 Cinder 进行评估，结果显示其漏洞显著少于其他流行的开源权重模型。有关我们安全方法的更多信息请参阅此处。

FLUX 3 Image Klein is the second open weights release from the FLUX 3 family, following FLUX 3 Action. More will follow.

FLUX 3 Image Klein 是 FLUX 3 系列的第二个开源权重版本，继 FLUX 3 Action 之后发布。更多版本即将推出。

Learn more

了解更多

  • FLUX 3: Real World Models, the research behind the FLUX 3 family
  • FLUX 3 Action, our open weights 7B model for robot control
  • FLUX 3：现实世界模型，FLUX 3 系列背后的研究
  • FLUX 3 Action，我们用于机器人控制的开源 7B 模型

Citation

引用

If you use FLUX 3 Image Klein in your work, please cite:

如果您在作品中使用了 FLUX 3 Image Klein，请引用：

代码 · 6 行
@misc{bfl2026flux3imageklein,
  author       = {{Black Forest Labs}},
  title        = {{FLUX} 3 Image Klein},
  year         = {2026},
  url          = {https://bfl.ai/blog/flux-3-image-klein},
}

Copy

复制

更进一步：量化金融体系

看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力

进入量化体系 →
话题#Black Forest Labs#FLUX 3 Image Klein#开源模型

相似阅读

关联信息，但可能不是同一事件