FLUX 3 Video Explained: Inside Black Forest Labs' Native-Audio Multimodal Model

Most AI video models still treat sound as an afterthought. You generate the clip, then you generate or bolt on audio separately, then you spend time syncing the two so a mouth movement lines up with a word or a footstep lands where a foot actually touches the ground. Black Forest Labs shipped something different on August 4, 2026. FLUX 3 Video generates video and audio from the same model, in the same pass, and it can also drive robot actions from the same learned representation [1].

That last part is not a marketing flourish. FLUX-mimic, a robotics system built on the FLUX 3 backbone, is already running on Audi's production floor doing manipulation work that conventional automation could not handle [7]. This post walks through what FLUX 3 Video actually generates, the research technique that makes one model good at video, audio, and action prediction at once, and how to call it yourself.

What Black Forest Labs Built

Black Forest Labs was founded in 2024 by Robin Rombach, Andreas Blattmann, and Patrick Esser, three former Stability AI researchers who had previously co-authored the papers behind latent diffusion, Stable Diffusion XL, and Stable Video Diffusion during their time at LMU Munich and Stability AI [11]. The company's first release, the FLUX.1 family of image models, quickly became one of the most widely used open and commercial text-to-image lineups. FLUX 3 is their first release to generate video, and the first to unify image, video, audio, and robot action prediction inside one architecture [2].

FLUX 3 Video and FLUX 3 Action shipped in early access alongside the main release, with FLUX 3 Image following in the following weeks. Access runs through the BFL API and a small number of launch partners, including fal, rather than a broad consumer rollout [6].

Step 1: One Flow-Matching Backbone, Not Separate Models Bolted Together

The technique underneath FLUX 3 is called Self-Flow, a self-supervised flow matching framework Black Forest Labs published in March 2026 in a paper titled "Self-Supervised Flow Matching for Scalable Multi-Modal Synthesis" [3]. Standard flow-matching video models produce realistic footage, but the internal features that drive generation are entangled in ways that make them hard for a separate system, like a robot action decoder, to read back out.

Self-Flow addresses that by training one backbone to denoise video, audio, and, when relevant, action tokens along the same flow-matching schedule, instead of relying on a separate feature-alignment model like DINOv2 to keep representations consistent across modalities. Removing that dependency simplifies the training stack and lets the model specialize further as compute and parameters scale, where older alignment methods tend to plateau [5]. Reported results show Self-Flow reaching the same training milestone in roughly 143,000 steps versus the step count needed by REPA, the prior industry-standard alignment method, a 2.8 times improvement in training efficiency [3].

Here is a simplified version of what that shared training objective is doing conceptually:

python
/code
import torch

def self_flow_loss(model, video_frames, audio_waveform, robot_actions=None):
    """Simplified illustration of Self-Flow's self-supervised flow matching
    objective: one shared backbone learns to denoise video, audio, and
    optionally robot action tokens along the same flow-matching schedule,
    instead of training a separate encoder like DINOv2 just to align
    features across modalities."""
    noise_video = torch.randn_like(video_frames)
    noise_audio = torch.randn_like(audio_waveform)
    t = torch.rand(1)

    noisy_video = (1 - t) * video_frames + t * noise_video
    noisy_audio = (1 - t) * audio_waveform + t * noise_audio

    pred_video, pred_audio, features = model(noisy_video, noisy_audio, t)
    video_loss = ((pred_video - (noise_video - video_frames)) ** 2).mean()
    audio_loss = ((pred_audio - (noise_audio - audio_waveform)) ** 2).mean()

    action_loss = torch.tensor(0.0)
    if robot_actions is not None:
        # FLUX-mimic reuses these same intermediate features to decode
        # actions, instead of training a separate vision-language-action
        # model from scratch.
        pred_actions = model.action_head(features)
        action_loss = ((pred_actions - robot_actions) ** 2).mean()

    return video_loss + audio_loss + action_loss

FLUX 3 was trained on tens of millions of hours of general video to learn broad world dynamics, plus hundreds of thousands of hours focused specifically on human and robot manipulation footage so the same backbone would be usable for FLUX-mimic later. Video prediction consumes more than 95 percent of the total training compute, with audio making up under 0.5 percent of training tokens, a reminder of how much of "generating audio well" actually comes from the shared video representation rather than a dedicated audio-only budget [3].

Step 2: What FLUX 3 Video Actually Outputs

FLUX 3 Video generates clips up to 20 seconds long, in any whole-number duration from 5 to 20 seconds, or left on auto for the model to decide, at 720p or 1080p resolution and 24 frames per second [9]. Audio comes out of the same flow-matching pass as the video frames. There is no separate diffusion model generating sound and no lip-sync post-processing step stitched on afterward. One generation call produces a complete audiovisual clip, dialogue, ambient sound, and lip movement included, in the language you write your prompt in, with support for English in multiple dialects plus Chinese, Spanish, French, German, Japanese, Portuguese, Russian, Italian, Indonesian, Turkish, Hindi, Punjabi, and more [1].

That single detail, audio generated inside the same pass as the video rather than layered on afterward, is the reason a spoken line in a FLUX 3 Video clip tends to land on the mouth movement without a separate alignment step. It is also the same underlying mechanism the Self-Flow paper is built around: one representation, multiple output modalities, decoded together instead of stitched together.

Step 3: Four Ways to Generate a Clip

FLUX 3 Video supports four distinct generation modes, and picking the right one matters more than most people expect.

Video continuation deserves a closer look, since it behaves differently from a typical outpainting feature. You feed the model up to four seconds of existing video and audio, tell it what should happen next, and it takes both the visual motion and the existing audio into account when extending the clip, keeping camera behavior, dialogue, and ambient sound consistent with what came before rather than restarting from scratch [1].

Multi-shot scene creation is the mode most relevant to short-form content specifically, since it generates several distinct shots as one coherent scene inside a single call, rather than requiring a separate generation and manual stitch for every camera angle. Here is a standalone prompt built to exercise that mode along with native dialogue audio:

Multi-shot 12 second commercial for a small coffee roastery. Shot 1, three seconds: close up of coffee beans pouring into a grinder, mechanical whirring sound. Shot 2, four seconds: a barista in a warm-lit shop pulls an espresso shot, steam hissing, camera slowly pushes in. Shot 3, five seconds: the barista hands the cup across the counter and says in English, here you go, careful it's hot, ambient cafe chatter and a bell above the door. Warm natural lighting throughout, consistent shop setting across all three shots, synchronized native audio, 1080p, no on-screen text.

Step 4: Calling FLUX 3 Video From Code

FLUX 3 is accessible through the BFL API directly, and through launch partners including fal, which bills output per second of generated video: $0.17 per second at 720p and $0.29 per second at 1080p on most endpoints, rising to $0.41 and $0.53 per second specifically on the Extend Video endpoint [10]. A basic text-to-video call through fal's Python client looks like this:

python
/code
import fal_client

result = fal_client.subscribe(
    "blackforestlabs/flux-3/text-to-video",
    arguments={
        "prompt": (
            "A rain-soaked neon alley in Tokyo at night, a street vendor calls "
            "out over sizzling food, camera pushes in slowly, ambient traffic "
            "and rain audio"
        ),
        "duration": 10,
        "resolution": "1080p",
    },
    with_logs=True,
)

print(result["video"]["url"])

Step 5: How FLUX 3 Benchmarks Against Other Video Models

Black Forest Labs reports FLUX 3 leading an all-vs-all text-to-video arena with an ELO rating of 1135, and tying Seedance 2.0 specifically on image-to-video quality, ahead of other existing video models tested in the same comparison [1]. Arena-style ELO comparisons like this should be read as a relative ranking among the specific models included in that evaluation round rather than an absolute quality score, but a leading all-vs-all result at launch is a meaningfully strong starting position for a first-generation video model from a lab that had never shipped a video model before.

Case Study: FLUX-mimic Puts the Same Backbone on a Factory Floor

The clearest evidence that FLUX 3's shared representation actually works across modalities is FLUX-mimic, built jointly by Black Forest Labs and robotics company mimic. FLUX-mimic decodes robot actions directly from FLUX 3's learned world representation, training a lightweight action decoder on the intermediate features already produced by FLUX 3's video prediction pathway, rather than training a separate vision-language-action model from scratch [7].

The system has been tested and deployed at Audi's production facilities, handling tasks conventional automation cannot manage well: kitting small parts into structured trays, inserting electronic control units, assembling components, and manipulating soft, flexible materials like seals and cables [7]. Christoph Schneider of Audi's Production Lab described it as solving complex soft-body manipulation work that would have been impossible with conventional robotics [7].

On performance, FLUX-mimic reports a reaction time of 101 milliseconds for self-contained robot systems, with backbone inference under 80 milliseconds on an NVIDIA RTX 5090, and outperforms prior vision-language-action models even when the FLUX backbone itself is kept frozen during training [7]. That last detail matters for the video generation side of FLUX 3 too. If a frozen video backbone already contains enough of a world model to drive real robot manipulation accurately, it is a strong signal that the video predictions themselves are grounded in physically plausible dynamics, not just visually convincing motion.

If you want to see the output quality in motion rather than read about it, this walkthrough covers real generated clips from the launch:

Why a Unified Backbone Matters Beyond the Benchmark Numbers

It is tempting to read FLUX 3 as just another video model with a good demo reel, but the architecture choice is the actual story. Treating video, audio, and action as three separate problems requiring three separate models is the default in most of the industry, and it shows up as friction everywhere: audio that almost syncs, robot policies trained from scratch on limited data, generation pipelines that need several models chained together to produce one usable clip. A single backbone trained to represent all three removes an entire category of integration work.

That same idea, fewer separate systems producing a more coherent result, is behind how a tool like Text2Shorts in Miraflow AI turns one topic into a finished script, visuals, and voiceover without the creator manually stitching several disconnected tools together. The cinematic AI video generator in Miraflow AI applies the same principle to turning a single prompt into a finished clip, and the AI Music Generator follows the same logic for audio specifically. If your workflow starts with a long video instead of a blank prompt, AI Clipping applies a related idea in reverse, analyzing a full upload once to find and caption the viral moments automatically rather than requiring a separate tool for transcription, scoring, and cropping.

We covered a similar audio-video tradeoff when comparing MiniMax H3 against Veo 3.1, and if native audio generation is what caught your attention here, our breakdown of Google's Lyria 3.5 covers the same shift happening on the music side. You can find more model breakdowns like this on the Miraflow AI blog.

Common Mistakes and Misunderstandings

A few things trip people up when they first work with FLUX 3 Video.

• Assuming longer prompts always produce better multi-shot results. FLUX 3 follows explicit shot-by-shot instructions well, but a prompt that tries to cram too many unrelated beats into 20 seconds still produces a rushed, incoherent clip. Structure shots the way you would structure an actual storyboard.
• Treating video continuation as simple outpainting. It is conditioned on both the existing video and its audio, so continuing a clip with silent footage and then expecting rich generated audio in the continuation works less reliably than feeding in audio from the start.
• Confusing the ELO benchmark with an absolute quality score. It reflects standing within a specific all-vs-all comparison at a point in time, not a fixed, universal ranking.
• Assuming Self-Flow is only a video technique. The paper's core claim is modality-agnostic self-supervised flow matching, which is exactly why the same backbone extends into audio and, through FLUX-mimic, into robot action prediction [3].

Production Notes for Working With FLUX 3 Video

• Budget per-second, not per-clip. At $0.17 to $0.29 per second depending on resolution, a 20 second 1080p clip costs meaningfully more than a 5 second one, so match duration to what the shot actually needs [10].
• Write dialogue in the language you want spoken, directly in the prompt, rather than describing the language separately, since lip-sync and pronunciation are generated together with the rest of the clip [1].
• For longer sequences, generate a strong first clip, then use video continuation with its existing audio intact, rather than generating several independent clips and trying to stitch them together afterward.
• If your use case involves physical products or repeatable actions rather than pure video content, keep an eye on FLUX-mimic. The same reasoning that makes FLUX 3 good at generating plausible video motion is the reasoning Black Forest Labs is using to justify it as a robotics foundation model too [7].

Frequently Asked Questions

Is FLUX 3 Video the same thing as FLUX.1, Black Forest Labs' image model? No. FLUX.1 is the company's earlier text-to-image lineup. FLUX 3 is a newer, separate multimodal model built on the Self-Flow research, and it is the company's first model capable of generating video and audio.

Can FLUX 3 Video generate spoken dialogue with lip sync? Yes. Audio, including spoken dialogue, is generated in the same pass as the video, across a wide list of supported languages, without a separate lip-sync step [1].

How long can a single FLUX 3 Video clip be? Up to 20 seconds per generation, in whole-second increments from 5 to 20, or left on auto. Video continuation lets you extend a clip further using up to 4 seconds of an existing clip as context [1] [9].

Is FLUX 3 open source? FLUX 3 Video launched through the BFL API and select partners, not as open weights. Black Forest Labs has released open weight versions of prior FLUX models before, but FLUX 3's initial video release is API access only.

What is FLUX-mimic and is it related to FLUX 3 Video? FLUX-mimic is a robotics system built on the same FLUX 3 backbone, developed with mimic robotics, that decodes robot actions from the model's learned video representation instead of training a separate action model from scratch. It is already deployed on Audi's production floor [7].

Do I need to understand Self-Flow to use FLUX 3 Video? No. Self-Flow is the training technique behind the model, not something you interact with directly through the API. Understanding it just explains why FLUX 3's video, audio, and action outputs behave as one coherent system instead of three stitched-together ones.

Conclusion

FLUX 3 Video is a genuinely different kind of release from the usual "better video model" launch. The interesting part is not just clip quality or the 1135 ELO score, it is that Black Forest Labs trained one backbone to represent video, audio, and physical action well enough that a lightweight decoder can pull real robot manipulation out of it. That is a strong argument that Self-Flow's shared representation is learning something closer to an actual world model than a purely visual pattern matcher. Whether you end up using FLUX 3 Video for a short-form ad, a dialogue-driven scene, or you are watching the robotics side more closely, the underlying idea, fewer separate models producing a more coherent result, is worth understanding now while it is still new.

References and Sources

[1] FLUX 3 Video, Part 1: Generation, Black Forest Labs.

[2] FLUX 3: Multimodal Video, Image and Audio, Black Forest Labs.

[3] Self-Supervised Flow Matching for Scalable Multi-Modal Synthesis, arXiv:2603.06507.

[4] Self-Flow code and project page, Black Forest Labs GitHub.

[5] Black Forest Labs' new Self-Flow technique makes training multimodal AI models 2.8x more efficient, VentureBeat.

[6] Black Forest Labs launches FLUX 3, VentureBeat.

[7] FLUX 3 x mimic: The Next Generation of Video-Action Models, Black Forest Labs.

[8] FLUX 3: Video with Native Audio by Black Forest Labs, fal.

[9] FLUX 3 Text to Video API docs, fal.

[10] fal Platform API pricing.

[11] Stable Diffusion creators launch Black Forest Labs, secure $31M for FLUX.1, VentureBeat.

[12] Black Forest Labs Releases FLUX 3, MarkTechPost.