FLUX 3 是在 FLUX.1 / FLUX.2（仅图像）之上扩展出的统一多模态 flow matching 基础模型，联合训练图像、视频、音频与动作预测。核心训练方法是 Self-Flow——一种把自监督特征重建目标嵌入 conditional flow matching 的框架，作者称其形成"更好的表示 → 更好的生成 → 更多训练信号 → 更好的表示"的正反馈循环。文中所有基准结果均标注为 BFL 训练中期的初步评测。FLUX 3 is a unified multimodal flow matching foundation model extended from FLUX.1 / FLUX.2 (image only), jointly trained on image, video, audio and action prediction. Its core training method is Self-Flow — a framework that embeds a self-supervised feature reconstruction objective into conditional flow matching; the authors describe it as forming a positive feedback loop of “better representations → better generation → more training signal → better representations”. All benchmark results in the article are marked as preliminary evaluations from the middle of BFL's training run.
此前的 FLUX.1 / FLUX.2 只覆盖图像生成。FLUX 3 的出发点是：图像、视频、音频与动作只是同一个物理现实的不同"投影"，把它们联合学习会互相约束，从而逼迫模型学到物理规律。The earlier FLUX.1 / FLUX.2 covered image generation only. FLUX 3 starts from the premise that image, video, audio and action are merely different “projections” of the same physical reality, so learning them jointly makes them constrain one another and forces the model to learn physical laws.
"images, video, audio, and actions are partial projections of the same underlying physical reality... the sound has to match the impact, the motion has to obey the mass, the future has to follow from the past."
这也解释了为什么视频被作为主导训练信号：视频占据 FLUX 3 训练算力的 >95%，因为视频天然要求模型学会接触、运动、重量与因果关系——任何一项出错都会在画面上直接暴露；相比之下音频是"简单模态"（720p 带音频视频里音频 token 占比 <0.5%），一旦模型学会了视觉世界动力学，把视觉事件映射到对应声音就容易得多。This also explains why video is used as the dominant training signal: video accounts for >95% of FLUX 3's training compute, because video inherently requires the model to learn contact, motion, weight and causality — an error in any one of them is immediately exposed on screen. Audio, by contrast, is a “simple modality” (audio tokens account for <0.5% of a 720p video with audio); once the model has learned visual world dynamics, mapping visual events to the corresponding sound is far easier.
FLUX 3 是一个联合训练于图像、视频、音频的 diffusion transformer（DiT），通过同一个多模态主干扩展出视频、原生音频、图像合成/编辑与动作预测能力。核心训练创新是 Self-Flow——由 Chefer、Esser 等人在 Black Forest Labs 与 MIT 提出的框架。FLUX 3 is a diffusion transformer (DiT) jointly trained on image, video and audio, extending video, native audio, image synthesis/editing and action prediction capabilities from one shared multimodal backbone. The core training innovation is Self-Flow — a framework proposed by Chefer, Esser and colleagues at Black Forest Labs and MIT.
标准 flow matching 通过预测"噪声到数据"的路径隐式学到一个世界模型，但其内部表示并不好地解耦——模型能生成一段球体下落的完美视频，却不一定把"重力"这个概念暴露为可提取的独立特征。Self-Flow 通过联合优化两个目标来解决这个问题：Standard flow matching implicitly learns a world model by predicting the “noise-to-data” path, but its internal representations are poorly disentangled — the model can generate a flawless video of a falling sphere without necessarily exposing “gravity” as an extractable, independent feature. Self-Flow addresses this by jointly optimizing two objectives:
据文中描述，在初步评测中 Self-Flow 同时提升了图像、视频、音频的生成质量（用归一化到"标准 flow matching = 100"的 Fréchet 距离衡量），并把微调后机器人操纵任务的成功率从 42% 提升到 71%（+29 个百分点）。文中把这个机制称为"reciprocal feedback loop"：更好的表示 → 更好的生成 → 更多训练信号 → 更好的表示。As described in the article, in preliminary evaluations Self-Flow improved generation quality for image, video and audio alike (measured by a Fréchet distance normalized so that “standard flow matching = 100”), and raised the success rate on robot manipulation tasks after fine-tuning from 42% to 71% (+29 percentage points). The article calls this mechanism a “reciprocal feedback loop”: better representations → better generation → more training signal → better representations.
FLUX 3 对物理动力学的理解延伸到了动作预测，有两条路径：Native action prediction——通过放大 Self-Flow 方法直接集成进 FLUX 3；以及Lightweight action decoder——在预训练视频预测路径提取的中间特征上训练，此路径由 mimic 更早的 mimic-video 工作率先提出，可以在不重训整个模型的前提下从主干学到的世界表示中解码机器人动作。FLUX 3's understanding of physical dynamics extends to action prediction along two paths: Native action prediction — integrated directly into FLUX 3 by scaling up the Self-Flow method; and Lightweight action decoder — trained on intermediate features extracted from the pretrained video prediction path, a route first proposed by mimic's earlier mimic-video work, which can decode robot actions from the world representation learned by the backbone without retraining the whole model.
基于此，Black Forest Labs 与 mimic robotics 联合构建了 FLUX-mimic——一个建立在 FLUX 3 主干上的 video-action 模型，已在 Audi 部署，用于此前传统机器人方案难以完成的软体材料操纵任务（密封件、线缆、柔性材料的装配与插入）。文中特别强调：即使冻结 FLUX 主干（frozen backbone），该动作解码器也能超越以往 VLA 模型（这是以往 VLA 模型完全失效的设定）；微调后的主干则在操纵基准上达到 state-of-the-art。作者还报告了一种未在训练数据中出现过的涌现行为——机器人抓取失败后会自我纠正，被解读为源自主干学到的物理理解。延迟方面，主干从输入到世界表示的处理在单张 RTX 5090 上 <80ms，加上传感器与执行器的全栈反应延迟为 101ms。Building on this, Black Forest Labs and mimic robotics jointly built FLUX-mimic — a video-action model built on the FLUX 3 backbone, already deployed at Audi for soft-material manipulation tasks that conventional robotics approaches struggled with (assembly and insertion of seals, cables and flexible materials). The article stresses that even with a frozen FLUX backbone, the action decoder surpasses previous VLA models (a setting in which previous VLA models fail completely); with a fine-tuned backbone it reaches state-of-the-art on manipulation benchmarks. The authors also report an emergent behavior never seen in the training data — the robot self-corrects after a failed grasp, interpreted as arising from the physical understanding learned by the backbone. On latency, the backbone's processing from input to world representation takes <80ms on a single RTX 5090, and the full-stack reaction latency including sensors and actuators is 101ms.
说明：官方公开材料未披露 FLUX 3 生产版本的完整实现细节（encoder / decoder / attention 布局 / 服务细节均未公开），也未公开参数量。Note: The official public materials do not disclose the full implementation details of the production version of FLUX 3 (encoder / decoder / attention layout / serving details are all undisclosed), nor the parameter count.
以下结果均为 BFL 在开发阶段的初步评测，原文注明"evaluations are early and we expect further improvements during the early access phase"，测试使用 720p、带音频的 10 秒文本转视频片段。All results below are preliminary evaluations by BFL during development; the original text notes that “evaluations are early and we expect further improvements during the early access phase”. Testing used 720p, 10-second text-to-video clips with audio.
| 对比对象Comparison | FLUX 3 偏好率FLUX 3 preference rate |
|---|---|
| vs. Luma Ray 3.2 | 93% |
| vs. Runway Gen-4.5 | 77% |
| vs. Grok Imagine Video | 69% |
| vs. Kling v3 Pro | 60% |
| vs. Happy Horse v1 | 59% |
| vs. Happy Horse 1.1 | 57% |
| vs. Seedance 2.0 | 52% |
| vs. Gemini Omni Flash | 52% |
BFL 强调 FLUX 3 在以下方面尤其突出：人脸表情与细微情绪、音画同步（声音与物理事件匹配）、多语言能力，以及使用视觉参考时跨多镜头的角色一致性。BFL stresses that FLUX 3 is especially strong at: facial expressions and subtle emotion, audio-visual synchronization (sound matching physical events), multilingual capability, and character consistency across multiple shots when using visual references.
| 指标Metric | 标准 Flow MatchingStandard Flow Matching | Self-Flow |
|---|---|---|
| 微调后机器人操纵成功率Robot manipulation success rate after fine-tuning | 42% | 71% |
| 达到目标成功率所需训练步数Training steps needed to reach the target success rate | Baseline | ~50%（约 2× 样本效率）~50% (about 2× sample efficiency) |
作为参照，mimic-video（前身方法）另外报告了相较 VLA 模型最高 10× 样本效率；FLUX-mimic 同时结合了这两方面的效率提升。Backbone 层面，Self-Flow 实验中 BFL 报告在同等算力下，用归一化的 Fréchet-distance 风格指标衡量，图像、视频、音频的生成质量相较标准 flow matching 均有提升，这些结果被用作把 Self-Flow 扩展到 FLUX 3 生产级训练的验证依据（原文未公开具体归一化数值）。For reference, mimic-video (the predecessor method) separately reports up to 10× sample efficiency relative to VLA models; FLUX-mimic combines both kinds of efficiency gain. At the backbone level, in the Self-Flow experiments BFL reports that under equal compute, measured by a normalized Fréchet-distance-style metric, generation quality for image, video and audio all improved over standard flow matching; these results serve as the validation basis for scaling Self-Flow to production-level FLUX 3 training (the original text does not disclose the specific normalized values).
| 能力Capability | 接入方式Access | 状态Status |
|---|---|---|
| FLUX 3 Video（视频+音频生成）FLUX 3 Video (video + audio generation) | API + 私有权重API + private weights | Early Access（已可用）Early Access (available now) |
| FLUX 3 Image（合成+编辑）FLUX 3 Image (synthesis + editing) | API + 私有权重API + private weights | Early Access（数周内上线）Early Access (live within weeks) |
| FLUX-mimic（动作预测）FLUX-mimic (action prediction) | 指定研究/商业合作伙伴Selected research / commercial partners | 已启动（始于 mimic robotics）Started (beginning with mimic robotics) |
| FLUX 3 Dev（开放权重主干）FLUX 3 Dev (open-weights backbone) | Open weights | 待后续发布To be released later |
所有基准结果都是训练中期的初步评测，BFL 明确表示 Early Access 阶段"further improvements"仍在预期之中，目前还没有社区独立验证过的基准结果。All benchmark results are preliminary evaluations from the middle of training; BFL states explicitly that “further improvements” are still expected during the Early Access phase, and there are as yet no independently community-verified benchmark results.
开放权重版本 FLUX 3 Dev 没有发布日期；在权重可用之前，API 之外的独立评测与微调实验都无法进行。The open-weights version FLUX 3 Dev has no release date; until the weights are available, independent evaluation and fine-tuning experiments outside the API are impossible.
20 秒相较此前的开放模型是显著提升，但仍低于部分专有系统的多分钟级生成能力。20 seconds is a marked improvement over earlier open models, but still falls short of the multi-minute generation capability of some proprietary systems.
FLUX 3 的许可条款尚未公布；作为参照，FLUX.2 Klein 是 Apache 2.0，FLUX.2 dev 使用非商业许可，社区仍需等待关于商用、微调权利与再分发的明确说明。The license terms for FLUX 3 have not been announced; for reference, FLUX.2 Klein is Apache 2.0 while FLUX.2 dev uses a non-commercial license, so the community must still wait for a clear statement on commercial use, fine-tuning rights and redistribution.
除了"数千万小时视频"这类高层次数量描述外，BFL 未公开训练数据的组成、过滤方法或来源，限制了可复现性研究。Beyond high-level quantity descriptions such as “tens of millions of hours of video”, BFL has not disclosed the composition, filtering methods or sources of the training data, which limits reproducibility research.
FLUX-mimic 目前只在单一工业合作伙伴（Audi）处针对特定操纵任务部署，向更广泛机器人领域的泛化能力尚未被表征。FLUX-mimic is currently deployed only at a single industrial partner (Audi) for specific manipulation tasks; its generalization to broader robotics domains has not been characterized.