FLUX 3: Black Forest Labs Ships One Model for Video, Audio and Robotics

Black Forest Labs (BFL) has released FLUX 3, a multimodal foundation model that generates images, video, and audio from a single set of weights. It is the first model in the FLUX family to cover video, audio, and robot action prediction together. In human preference tests using 10-second, 720p text-to-video clips with audio, FLUX 3 was preferred over Runway Gen-4.5 in 77% of comparisons and over Luma Ray 3.2 in 93%, though it lands at roughly 52% against Seedance 2.0 and Gemini Omni Flash.

What happened

Detail	Fact
Developer	Black Forest Labs (BFL)
Model name	FLUX 3
Max video length	20 seconds per generation
Test resolution	720p with native audio
Underlying method	Self-Flow (introduced March 2026)
Reference implementation license	Apache 2.0
Base architecture	SiT-XL/2 with per-token timestep conditioning

FLUX 3 is built on a framework BFL calls Self-Flow, which combines a flow matching training objective with a self-supervised feature reconstruction objective. The reference implementation on GitHub trains with a 25% per-token mask ratio and uses self-distillation from an EMA (exponential moving average) teacher at layer 20 to a student at layer 8. BFL launched Self-Flow in March 2026, so the method itself is not new. What is new with FLUX 3 is the scale: BFL says it “significantly scaled up compute and data resources” to train across video, images and audio simultaneously.

The core argument from BFL is that images, video, and audio are each a “lossy projection” of the same underlying reality. Training all three modalities together forces them to constrain one another: a sound must match the physical impact that causes it, and motion must respect the physics of mass. The team says that training on a single modality therefore gives an incomplete picture of the world.

What FLUX 3 Video can do

The video generation side covers five distinct modes:

• Text-to-video
• Image-to-video
• Video-to-video from a reference clip
• Keyframe-to-video for controlled transitions
• Generative video-audio continuation from input video and audio

BFL also highlights multilingual dialogue generation, agentic chaining of clips into multi-shot sequences, strong typography with animated designs, and particularly good handling of human facial expressions. The audio is native, not bolted on after the fact.

How does FLUX 3 compare to competitors?

BFL published preliminary human preference results. Evaluators compared 10-second text-to-video clips at 720p with audio. The table below shows what share of comparisons FLUX 3 won.

Competitor model	FLUX 3 preference rate
Luma Ray 3.2	93%
Runway Gen-4.5	77%
Grok Imagine Video	69%
Kling v3 Pro	60%
Happy Horse v1	59%
Happy Horse 1.1	57%
Seedance 2.0	52%
Gemini Omni Flash	52%

The 93% and 77% numbers are strong, but the 52% results against Seedance 2.0 and Gemini Omni Flash are essentially a coin flip. BFL calls these “preliminary” results, so independent benchmarking has not yet confirmed the figures.

Why it matters

For creative and marketing teams, native audio in a single generation pass removes a painful post-production step. Matching sound to motion has historically required a separate tool and a separate budget. A model that ties physical plausibility into its training signal could produce more convincing results without manual audio syncing.

The robot action prediction angle is less immediately useful for most digital teams but signals where the model is headed. Training on physical cause-and-effect is relevant for any application that needs a model to reason about how the world actually works, not just how it looks. That could matter for product visualisation, interactive demos, or the kind of AI integration work that involves connecting vision models to downstream systems.

For teams already evaluating AI video tools for ad production or social content, the 720p native-audio clips are directly relevant. If the preference numbers hold up under independent testing, FLUX 3 would sit above Runway Gen-4.5 and Luma Ray 3.2 in quality, which are currently two of the most widely used options. We have been watching this space closely in our AI news coverage, including the recent Runway deal with Netflix and Lionsgate.

Our take

The BFL results are self-reported and the evaluation setup (human preference on a curated set of prompts) favours the team running the test. That is a standard move at launch and it does not mean the numbers are wrong, but treat them as a ceiling until independent reviews land.

The genuine news here is the architecture decision: one model, one set of weights, four output types. That matters more than any single benchmark score because it changes the operational cost of running the model in production. Fewer specialised models to maintain, fewer API calls to different endpoints, and a single fine-tune path if a client wants custom output. For agencies and product teams building automated content workflows, that simplification is real money.

The robot action prediction output is worth watching separately. BFL framing this as a general multimodal model rather than a video tool suggests they are building toward something bigger. Whether that turns into accessible developer tooling or stays a research capability for the next year is the open question.

If you are evaluating AI video for ad production or social content and want to know whether FLUX 3 fits your workflow, get in touch with the Lumien team for a straight answer based on your actual use case.

Practical takeaway: Wait for independent benchmarks before switching production workflows to FLUX 3, but put it on your test list now, especially if native audio sync is a current pain point.

Source: Marktechpost