How to Prompt FLUX 3 Video: The 4-Part Formula and Workflow

Short answer: Write every FLUX 3 video prompt in four parts — camera, scene, action, audio — because the model generates sound in the same pass as the picture. Set aspect ratio, frame rate, duration and reference weight before you generate, chain shots to go past 20 seconds, and check every clip on four points before it goes into the edit.

What is confirmed

Black Forest Labs — founded by the researchers behind VQGAN, latent diffusion and Stable Diffusion — calls FLUX 3 an omnimodel: one network trained across images, video, native audio and physical action, instead of separate systems for picture, voice and lip-sync. It hasn’t disclosed the parameter count. It has published the training method, Self-Flow, which in BFL’s paper reached text-to-image convergence about 2.8 times faster than REPA.

Animate a starting image.
Continue an existing video and audio clip.
Carry a character from a reference video into a new scene.
Up to 20 seconds per generation, with audio generated in the same pass.

At the time of the video, FLUX 3 video was in early access, with an open-weight dev version planned. For its status in Promptus, see FLUX 3 on Promptus.

The benchmark numbers, with the caveat

Compared with	FLUX 3 preferred
Luma Ray 3.2	93%
Runway Gen-4.5	77%
Grok Imagine Video	69%
Kling v3 Pro	60%
The rest of the field	52–59%

These are BFL’s own preliminary blind preference tests on 10-second 720p clips — company-run and not independently verified.

The 4-part prompt formula

Part	What to write
1. Camera	Trajectory, focal length, tracking speed
2. Scene and lighting	What’s in frame, composition, depth of field
3. Action and dynamics	What physically happens, including facial expression
4. Native audio	Dialogue, effects and ambience

Worked example from the video: Camera: a static wide shot with a sharp central vertical split. Scene: a home desk with a laptop glowing under warm light on the left; an enterprise server room with blue LED racks on the right. Action: a developer typing steadily on the left; cooling fans spinning up and network lights pulsing on the right. Audio: quiet keyboard clacks on the left channel blending into server fan hum and data chirps on the right.

Four settings before you generate

Setting	Choice
Aspect ratio	16:9 for widescreen, 9:16 for Shorts
Frame rate	24 fps for a filmic feel, 30 fps for crisper motion
Duration	Up to 20 seconds per pass
Reference weight	Start around 0.7; raise towards 0.95 only when consistency matters more than motion

The 4-phase production workflow

Pre-production: break the script into shots, make a keyframe for each, and give every recurring subject and location a reference.
Prompt consistency: keep camera, scene, lighting and reference guidance the same across every prompt.
Reference weight pass: start at 0.7, raise only where needed, and keep the seed fixed.
Chaining: for anything over 20 seconds, feed the last frame of shot one in as the first frame of shot two, and pass the whole previous clip as a secondary reference.

The pre-publish check

Check	If it fails
Temporal consistency (morphing faces, flicker)	Raise the reference weight
Physical realism	Lower motion intensity or add a reference clip
Audio sync	Rewrite the audio directive and keep the seed fixed
Prompt adherence	Re-check all four prompt parts

Don’t move a shot into the edit until it passes all four. Official source: Black Forest Labs: FLUX 3. For a local option with sound today, see five local image-to-video models tested.

Frequently Asked Questions

What is FLUX 3?

Black Forest Labs’ omnimodel: one network trained on images, video, native audio and physical action. FLUX 3 video takes text, images, reference video and audio as input and outputs up to 20 seconds of video with synchronised audio.

How should I write a FLUX 3 prompt?

In four parts: camera behaviour, scene and lighting, action and dynamics, and native audio. Skipping a part is where quality drops off.

What reference weight should I use in FLUX 3?

Start around 0.7 and raise it towards 0.95 only when consistency matters more than freedom of motion.

How do I make a FLUX 3 video longer than 20 seconds?

Chain shots: feed the final frame of one shot in as the starting frame of the next, and pass the previous clip as a secondary reference to keep motion, lighting and sound consistent.

Are the FLUX 3 benchmark numbers independent?

No. The preference results, such as 77% against Runway Gen-4.5, come from Black Forest Labs’ own preliminary early-access comparisons and have not been independently verified.

Written by:

Claudia Perez

Claudia Perez is a tech creator and founder at Promptus. She loves to design and share expressive image workflows. Her work focuses on visual storytelling, refined aesthetics, and ComfyUI workflows.

Try Promptus Cosy UI today for free.