Black Forest Labs 发布 Flux 3：可生成带原生音频的 20 秒视频，并推出机器人动作模型 Flux-mimic

德国人工智能公司黑森林实验室（Black Forest Labs, BFL）发布了Flux 3，一种多模态基础模型，可同时从图像、视频和音频中学习。在BFL的早期测试中，它在视频生成方面击败了多个竞争对手。

BFL将Flux 3描述为迈向“现实世界视觉智能”的一步，它将视觉智能定义为能够“在物理和数字环境中感知、预测和行动”的模型。该公司是构建所谓世界模型的更广泛推动的一部分：https://the-decoder.com/google-deepmind-argues-video-generators-already-contain-the-world-models-computer-vision-has-been-missing/。

BFL认为，没有单一模态能够完整捕捉现实。图像展示空间结构，视频捕捉其随时间的变化，音频可以揭示机械事件与其产生的声音之间的联系。对三者进行联合训练可以让它们互相填补空白，赋予模型比单独训练每种模态更多的信息。Ad

Flux 3现在可以首次生成带有原生音频的视频：https://bfl.ai/blog/flux-3，视频片段最长可达20秒。它支持文本到视频、图像到视频、视频到视频、基于关键帧的过渡、多语言对话，以及用于更长多镜头序列的由代理驱动的片段链接。BFL表示，该模型在处理人类面部表情和将声音与物理事件匹配方面特别出色。Ad DEC_D_Incontent-1

在使用720p、10秒视频片段的早期评估中，BFL报告称Flux 3在93%的比较中优于Luma Ray 3.2，在77%的比较中优于Runway Gen-4.5，在69%的比较中优于Grok Imagine Video。面对更强的竞争者时优势缩小。Flux 3在60%的时间内优于Kling v3 Pro，59%优于Happy Horse v1，57%优于Happy Horse 1.1，以及52%优于Seedance 2.0和Gemini Omni Flash。

BFL表示，这些结果为初步结果，目前尚无独立测试。若能匹配已经进入好莱坞的领先系统如Seedance：https://the-decoder.com/hollywood-wants-seedance-banned-and-reportedly-also-wants-to-keep-using-it/，以及Gemini Omni Flash，将使Flux 3跻身顶级视频模型行列。Ad

BFL还预计Flux 3将在图像生成方面有所提高，特别是在处理复杂提示和多语言文本准确呈现方面。公司计划在未来几周内推出Flux 3 Image的抢先体验版本。

BFL表示，该模型还可以基于对世界的理解预测动作。公司与Mimic Robotics合作开发了Flux-mimic，一种视频动作模型：https://bfl.ai/blog/flux-3-mimic，目前正在奥迪的生产任务中进行测试。广告 DEC_D_Incontent-2

Flux 3基于Self-Flow：https://bfl.ai/research/self-flow，这是BFL用于让一个模型同时生成和理解内容的方法。多模态变换器使用专门组件将图像、视频和音频转换为共享的内部表示，然后再转换回输出结果。广告

动作组件为机器人应用提供了基础。BFL表示，这种统一学习过程比之前标准的流匹配方法在生成质量和模型对物理世界的理解方面都能提供更好的结果。

BFL正在分阶段推出所有功能，并通过早期访问阶段进行反馈和安全测试。Flux 3 Video：https://bfl.ai/models/flux-3已可用，Flux 3 Image将在未来几周内推出。动作预测最初将通过精选合作伙伴提供。

BFL还计划以“Flux 3 Dev”的名义发布多模态骨干网络的开放权重访问。长期来看，公司正在研发下一代模型，旨在将感知、动作和语言预测结合于单一模型中。

保持对AI的关注。清晰、有用，无废话。

关注The Decoder以获取AI新闻、背景故事和专家分析。

The Decoder：https://the-decoder.com/

German AI company Black Forest Labs (BFL) has released Flux 3, a multimodal foundation model that learns from images, video, and audio together. In BFL's early tests, it beat several rivals in video generation.

BFL describes Flux 3 as a step toward "real-world visual intelligence," which it defines as models that can "perceive, predict, and act across physical and digital environments." The company is part of a broader push to build so-called world models：https://the-decoder.com/google-deepmind-argues-video-generators-already-contain-the-world-models-computer-vision-has-been-missing/.

No single modality captures reality in full, BFL argues. Images show spatial structure, video captures how it changes over time, and audio can reveal links between mechanical events and the sounds they produce. Training on all three together lets them fill gaps for one another, giving the model more information than training on each modality separately. Ad

Flux 3 can now generate videos with native audio：https://bfl.ai/blog/flux-3 for the first time, with clips up to 20 seconds long. It supports text-to-video, image-to-video, video-to-video, keyframe-based transitions, multilingual dialogue, and agent-driven links between clips for longer multi-shot sequences. BFL says the model is especially good at human facial expressions and matching sounds to physical events. Ad DEC_D_Incontent-1

In early evaluations using 10-second clips at 720p, BFL reports that Flux 3 was preferred over Luma Ray 3.2 in 93 percent of comparisons, over Runway Gen-4.5 in 77 percent, and over Grok Imagine Video in 69 percent. The margins narrow against stronger competitors. Flux 3 was preferred over Kling v3 Pro 60 percent of the time, over Happy Horse v1 at 59 percent, over Happy Horse 1.1 at 57 percent, and over both Seedance 2.0 and Gemini Omni Flash at 52 percent each.

BFL says the results are preliminary, and no independent tests are available yet. Matching leading systems such as Seedance, which has already reached Hollywood：https://the-decoder.com/hollywood-wants-seedance-banned-and-reportedly-also-wants-to-keep-using-it/, and Gemini Omni Flash would put Flux 3 among the top video models. Ad

BFL also expects Flux 3 to improve image generation, especially for complex prompts and accurate text rendering in multiple languages. The company plans to release Flux 3 Image in early access within the next few weeks.

BFL says the model can also predict actions based on its understanding of the world. The company worked with Mimic Robotics to develop Flux-mimic, a video-action model：https://bfl.ai/blog/flux-3-mimic now being tested on production tasks at Audi. Ad DEC_D_Incontent-2

Flux 3 is based on Self-Flow：https://bfl.ai/research/self-flow, BFL's approach for teaching one model to generate and understand content at the same time. A multimodal transformer uses dedicated components to convert images, video, and audio into a shared internal representation and then turn it back into outputs. Ad

A component for actions provides the foundation for robotics applications. BFL says this unified learning process delivers better results than the previously standard flow-matching method, both in generation quality and in the model's grasp of the physical world.

BFL is rolling out all capabilities in stages, with early-access phases for feedback and safety testing. Flux 3 Video is：https://bfl.ai/models/flux-3 already available, and Flux 3 Image is set to follow in the coming weeks. Action prediction will initially be offered through select partners.

BFL also plans to release open-weight access to the multimodal backbone under the name "Flux 3 Dev." Longer term, the company is working on next-generation models that aim to combine perception, action, and language prediction in a single model.

Stay in the loop on AI. Clear, useful, no fluff.

Follow The Decoder for AI news, background stories and expert analyses.

The Decoder：https://the-decoder.com/