> For the complete documentation index, see [llms.txt](https://docs.pletor.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.pletor.ai/models/video-models/flux-3.md).

# Flux 3

Black Forest Labs' first video model. Direct contender to Seedance, Kling and Veo families.

## Overview

Black Forest Labs' first cinematic video model: 20-second single-pass clips with native audio, frame-accurate lip-sync in 13+ languages, and keyframe control inside a single generation.

The step change is not resolution or motion quality alone. Picture, sound, dialogue, and camera choreography resolve in one pass. No separate audio model, no sync step.

> **Note:** Flux 3 takes keyframes, not references. There is no subject or style reference input — control comes from frames placed on a timeline. If you need reference-driven consistency, use Seedance 2.5 or Seedance 2.0.

***

#### **Strengths**

* **20-second one-takes**: Single generation, no cuts, no stitching. Subjects and geometry hold across large temporal jumps (day to night, season turn). Clips can be chained past 20s.
* **Native audio in the same render**: Sound is not a second model and not a post-sync step. Events arrive carrying their own sound on the exact frame.
* **Multilingual dialogue with lip-sync**: 13+ languages including English (multiple dialects), French, Spanish, Chinese, Japanese, German, Portuguese, Russian, Italian, Indonesian, Turkish, Hindi, and Punjabi. Mixed-language scenes hold in one take.
* **Keyframes as a control surface**: Set a start image, an end frame, or multiple keyframes inside a clip. The model interpolates between them while holding the intended visual language.
* **Camera behavior**: Orbits, dolly moves, focus racks, and tracking shots preserve parallax and scene geometry through continuous movement.
* **Aesthetic range**: Distinctive spike on nostalgic, retro, and vintage registers — 16mm film grain, archival looks, period-accurate texture.

#### **Ideal use cases**

* **Cinematic ads**: TV-grade spots with picture, score, and dialogue delivered as one asset.
* **Stop-scrolling creative**: Viral-register content in the Sora 2 family, with more directorial control.
* **Multilingual campaigns**: One master scene, localized dialogue with accurate lip-sync per market.
* **Frame-locked production**: When specific moments must land exactly, set them as keyframes and let the model resolve the space between.

#### **Weaknesses**

* **Keyframes, not references**: Flux 3 takes frames placed on a timeline rather than reference images for subjects or style. Consistency across separate generations requires reusing keyframes, not references.
* **Render time is a few minutes per generation**: Rules it out of fast iteration loops. Explore in a quick model (Gemini Omni Flash, Grok Imagine), then commit the finished idea to Flux 3.
* **Longer clips reward longer prompts**: 20 seconds means more decisions, and the model makes the ones you leave open. Underspecified prompts drift.

***

## How to use effectively

**Word order is weight.** The model reads the prompt front to back; put the subject and core action first, style and atmosphere after.

**Declare absences explicitly.** There are no negative prompts. Instead of "no music", write what is present: "only ambient street sound and footsteps."

**Quote dialogue verbatim, with a visible speaker.** Name who talks, put the line in quotes, state the language and accent if not obvious.

**Layer audio with verbs, not adjectives.** "Rain hits the awning, a bus hisses past" beats "moody urban soundscape."

**Keyframes carry the moments that must land.** Set the frames that have to be exact; size the space between them to the action and let the model pace itself.

**Use camera language.** Locked-off, push-in, pull-back, orbit, tracking — the model executes professional cinematography vocabulary literally.