Flux 3 Video Is Live: What Black Forest Labs' Open-Weight Bet Means
Flux 3 Video is now live in Runway and Leonardo, with an open-weight release promised. Here's how it compares to ByteDance's SeaDance 2.5.
What is Flux 3 Video?
Flux 3 Video is Black Forest Labs’ new AI video generation model, now accessible through platforms like Runway and Leonardo after being teased a few weeks earlier. It can generate clips up to 20 seconds long, and Black Forest Labs describes it as a model built to model reality as accurately as possible, suggesting a focus on physical plausibility rather than just visual polish. What makes it stand out isn’t raw realism. It’s the company’s stated plan to eventually release an open-weight version, something no major closed video model competitor has committed to at this scale.
TL;DR
- Flux 3 Video launched this week across Runway and Leonardo, generating clips up to 20 seconds long with a stated emphasis on realistic physics over flashy visuals.
- Side-by-side testing with identical prompts showed Flux 3 Video producing multi-angle, multi-scene outputs from a single prompt, switching camera angles and perspectives without being told to.
- Visually, Flux 3 Video reads as more obviously synthetic than ByteDance’s rival release, SeaDance 2.5, which launched the same week and leans toward photorealism.
- Black Forest Labs has confirmed plans for an open-weight release of Flux 3 Video, though details on whether it will run on consumer GPUs remain unannounced.
- SeaDance 2.5 supports up to 30 seconds of video and accepts a mix of 30 images, 10 video clips, and 10 audio clips as reference material in a single generation, plus timestamp-level editing control.
- Both models are currently usable through mainstream platforms (SeaDance via Dreamina/CapCut, Flux 3 Video via Runway and Leonardo), lowering the barrier for anyone wanting to test them directly.
- The gap between these tools and something like early AI video generation from just a few years ago is stark enough that distinguishing real footage from generated footage is becoming genuinely difficult.
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
How does Flux 3 Video actually perform?
In direct testing using identical prompts, such as a colorful underwater scene with alien sea life, and a portrait photo paired with a nightclub dance instruction, Flux 3 Video produced results that looked more like generative art than convincing realism. The company’s own example clips on its site don’t hide this either. Nobody is going to mistake Flux 3 Video’s output for real footage in its current form.
What’s notable is behavior, not polish. Given a single prompt, Flux 3 Video generated multiple distinct scenes and camera angles rather than one static shot. The underwater test showed a diver swimming toward a jellyfish, then a lower camera angle, then a cut to an alien-looking creature, then another angle change back to the diver, all from one text description with no explicit shot-list. That kind of autonomous scene-splitting is a meaningfully different capability from a model that just extends a single continuous shot.
How does it compare to SeaDance 2.5?
ByteDance released SeaDance 2.5 in the same week, and the two models make for a useful contrast because they were tested with the exact same prompts.
SeaDance 2.5 supports longer clips (up to 30 seconds versus Flux 3 Video’s 20), and it accepts a much richer set of reference inputs: up to 30 images, 10 video clips, and 10 audio clips in a single pass, along with timestamp-level control for editing specific moments in the audio or video. That’s a heavier, more production-oriented toolset aimed at people who want granular control over a longer piece of content.
Visually, SeaDance 2.5 also leans more photorealistic. Its demo outputs are convincing enough that the difference from real footage isn’t always obvious, though testing surfaced coherence problems, like a jellyfish that morphs into what looks like crystal formations when the camera angle changes. Flux 3 Video, by comparison, looks more overtly synthetic across the board but handles multi-angle scene composition in a way that felt more dynamic in testing.
Access differs too. SeaDance 2.5 is currently available through Dreamina (CapCut’s platform) and is not cheap to run in volume. Flux 3 Video is already integrated into multiple third-party platforms, including Runway and Leonardo, making it easier to try without committing to a single vendor’s pricing.
Why does Black Forest Labs’ open-weight plan matter?
Buried in Black Forest Labs’ announcement is a commitment to release an open-weight variant of Flux 3 Video. That’s unusual in the current video-model market, where the highest-performing systems, including SeaDance 2.5, are closed and accessible only through hosted platforms.
An open-weight release would let developers and researchers download the model, inspect it, fine-tune it, and integrate it into custom pipelines instead of depending entirely on a hosted API. That matters for anyone building products on top of video generation, since closed models can change pricing, usage limits, or availability without warning, while open weights offer more control and stability once you have them.
The unresolved question is scale. It is not yet clear whether the open-weight version will be small enough to run on consumer-grade GPUs, or whether it will require the kind of heavy cloud infrastructure needed for large open-weight language models. Until Black Forest Labs publishes concrete specs, hardware requirements, and licensing terms, it’s not possible to say how accessible this will really be for individual developers versus well-funded teams.
Is Flux 3 Video worth using right now?
For anyone doing serious realism-focused work, like marketing footage or content meant to look filmed, SeaDance 2.5 currently produces more convincing results out of the box. But Flux 3 Video’s multi-angle scene generation is a genuinely useful behavior for people building short narrative content, since it reduces the need to manually prompt for every camera cut.
The bigger reason to pay attention to Flux 3 Video isn’t this week’s output quality. It’s the open-weight roadmap. If Black Forest Labs follows through with a usable, reasonably sized open release, it would give developers an alternative to fully closed systems in a category where that option barely exists today.
Why does better AI video make misinformation harder to catch?
As these models improve, distinguishing real video from generated video is getting harder, not easier. Some of that improvement is being used well, for things like stylized trailers, edited sequences, and creative short-form content where the AI origin is either obvious or beside the point. But the same capability also feeds a growing volume of low-effort, misleading, or outright deceptive content online.
Watermarking, which some of these models already apply automatically, helps but isn’t a complete fix, since watermarks can be cropped, re-encoded, or otherwise stripped. The practical shift underway is less about spotting fakes with confidence and more about defaulting to skepticism: treating unverified video, especially anything unusually dramatic or convenient, as unconfirmed until there’s a reason to trust it.
Frequently Asked Questions
What is the maximum video length for Flux 3 Video?
Flux 3 Video can generate clips up to 20 seconds long, based on Black Forest Labs’ current release.
Will Flux 3 Video be open source?
Black Forest Labs has stated it plans to release an open-weight version of Flux 3 Video, though hardware requirements and timing haven’t been detailed publicly yet.
How is Flux 3 Video different from SeaDance 2.5?
SeaDance 2.5 supports longer clips (up to 30 seconds), accepts a wider range of reference inputs (images, video, and audio clips), and leans toward photorealism. Flux 3 Video caps at 20 seconds, looks more visibly synthetic, but generates multiple camera angles and scene cuts from a single prompt.
Where can I try Flux 3 Video?
Flux 3 Video is available through platforms including Runway and Leonardo, alongside other video generation tools on those services.
Is AI-generated video hard to detect now?
Yes, increasingly so. Some platforms add visible watermarks, but these can be removed or cropped, and the underlying video quality has advanced to the point where casual viewers often can’t reliably tell generated footage from real footage.
