Black Forest Labs' FLUX 3, Beyond Text-to-Image

FLUX 3 Video is Black Forest Labs' first capability from its new FLUX 3 family: a multimodal model that generates video and synchronized audio in one pass, across text-to-video, image-to-video, keyframe control, and video continuation. It's live now on Runware, through the API and Playground.

This article covers what makes FLUX 3 Video different, the new territory it opens up for BFL as a flexible video model, what you can actually create with it, and where it fits into a production workflow, plus ready-to-run API examples, and pricing to get started.

Black Forest Labs has dominated the text-to-image ecosystem since FLUX.1 launched in 2024, and FLUX.2 extended that lead. FLUX 3 Video is BFL's first step into video: the same architecture extended to train jointly across video, audio, image, and action prediction.

FLUX 3 Video: creative freedom across every style

What stood out most in testing is that it isn't built around one visual aesthetic. We ran it against realistic product footage, cinematic sequences, animation, motion graphics, and some fairly experimental prompts, and it held its ground on all of them.

Black Forest Labs' own numbers back that up: an ELO rating of 1135 leads the all-vs-all ranking for text-to-video, and in image-to-video it ties Seedance 2.0 while beating every other model tested, per Black Forest Labs.

What you can create with FLUX 3 Video

FLUX 3 Video handles both simple prompts and detailed, multi-element scenes, turning descriptions into coherent video, whether that's an advertisement, a character sequence, a product concept, or an entirely fictional world.

It also renders natural typography and written elements within scenes, so title cards, signage, and product packaging can be part of the shot rather than added afterward.

Start from a single reference image and animate forward from it. FLUX 3 Video keeps the source frame's look intact while your prompt drives the motion and camera work, with synced audio applied natively.

Pin frames in specific locations, such as where a shot should begin or end, and FLUX 3 Video generates the motion that connects them.

It functions more like a storyboard than prompting: define the key moments, and the model handles the transitions between them, synced audio included. Two keyframes with a fixed duration morph directly between them; two keyframes with duration: "auto", or three or more keyframes at any duration, generate a full storyboarded sequence instead.

Extend an existing clip from its final frames, carrying the same subjects, motion, style, and even audio forward into new footage.

Feed a previous generation back in as the starting point, and FLUX 3 Video continues from where it left off, drawing on up to 4 seconds of the existing clip's video and audio as context, useful for chaining generations into longer sequences or stretching a clip to a different runtime, without rebuilding the scene from scratch. Source clips are capped at 50MB and 15 seconds, and continue most cleanly from an ending where the motion has resolved.

A single generation can contain multiple connected scenes and camera angles, not just one isolated shot. Label shots directly in the prompt with SHOT 1:, SHOT 2:, and so on, separated by HARD CUT., and the model renders real in-video cuts rather than separate clips stitched together.

Dialogue, sound effects, and ambient audio all come out of the same pass as the video, so they match the scene rather than being bolted on afterward. Lip sync works across languages too: English (in various dialects), Chinese, Spanish, French, German, Japanese, Portuguese, Russian, Italian, Indonesian, Turkish, Hindi, Punjabi, and more.

To get spoken dialogue rather than burned-in captions, a prompt needs three things: a verbatim quoted line, a visible speaker, and an explicit instruction not to render on-screen text.

Text has historically been one of the harder things for video models to render cleanly. FLUX 3 Video handles it well enough that signs, packaging, title sequences, and motion graphics can be part of the creative direction rather than an afterthought.

FLUX 3 Video renders hyper-realistic footage, cinematic scenes, stylized animation, and motion design, all from the same model, so a project can shift look without switching model.

Chaining generations into longer workflows

Through agentic chaining and video continuation, generations connect into longer, automated pipelines: feed each clip's final frames into the next to build a sequence longer than any single generation, or hand the output off to another step in your pipeline.

Why Runware for FLUX 3 Video

FLUX 3 Video runs on Runware's pay-as-you-go multi-modal inference API alongside every other model on the platform, video, image, audio, and LLMs, under one key and one bill, with no dedicated infrastructure to provision or manage. There's no lock-in: swap between FLUX 3 Video, other video models, or a different provider entirely without rewriting your integration. Pricing follows Runware's standard usage-based model, billed per second of output (see the pricing page and the rates below).

Built for production on Runware

FLUX 3 Video is available now through Runware's multi-modal inference API and Playground. Full parameter documentation lives on the model's docs page. A request to https://api.runware.ai/v1 might look like:

{
"taskType": "videoInference",
"taskUUID": "{{$guid}}",
"model": "bfl:flux@3-video",
"positivePrompt": "A cinematic tracking shot through a neon-lit night market, steam rising from food stalls, ambient chatter and sizzling sounds",
"width": 1920,
"height": 1088,
"duration": 8,
"settings": {
  "audio": true
},
"outputFormat": "MP4",
"deliveryMethod": "async"
}

A few of the parameters worth knowing before your first request:

Pricing is usage-based, billed per second of output:

Guides

For hands-on technique, Runware's docs have five dedicated FLUX 3 Video guides:

• Prompting FLUX 3 — request shape, camera language, and how draft mode fits into an iteration loop
• Keyframes with FLUX 3 — pinning images to frame positions, from a single opening shot to full storyboards
• Video continuation with FLUX 3 — chaining generations cleanly, and recovering a shot that didn't land
• Multi-shot sequences with FLUX 3 — HARD CUT syntax, shot pacing, and threading audio across cuts
• Audio and speech with FLUX 3 — directing music, sound effects, and dialogue delivery

One model family across video, audio, image, and action

Generative AI models have traditionally been built around individual capabilities: one model for images, another for video, another for audio. FLUX 3 trains across all of them jointly, built on what BFL calls Self-Flow, their approach for aligning multimodal generation and understanding within one architecture. It develops a single understanding of how scenes behave: the relationships between visual information, movement, sound, and physics.

For creators, this means audio, motion, and visual detail are generated together in one pass. For developers and researchers, the same architecture extends to action prediction: the model can translate visual observations and instructions into predicted physical outcomes. BFL has already put this to work outside content creation entirely, partnering with mimic robotics to build FLUX-mimic, a video-action model combining the FLUX 3 backbone with robot-learning expertise for dexterous manipulation, now being tested on real production tasks at Audi.

Frequently asked questions