FLUX 3
A new multimodal model that makes up to 20-second videos, with sound, from a single prompt.
Jump to the full step-by-step walkthrough ↓
Quick demo
What it is
FLUX 3 is the newest model family from Black Forest Labs, the German AI lab behind the popular FLUX image models. FLUX 3 Video turns text or images into clips of up to 20 seconds with synchronized audio (dialogue, ambience, effects). The same model family is planned to cover image generation and editing, and later an open-weight Dev version.
Why people use it
- Longer clips than most video models, so you can do dialogue and continuous camera moves without stitching.
- Native audio means fewer extra tools in the workflow.
- Reported launch pricing is per second of video, with a cheap Draft mode for testing ideas before paying for a final render.
How to get started
- Go to bfl.ai and look for FLUX 3 Video (it is offered through their site and API, and some partner platforms).
- Start in Draft mode to test your prompt cheaply, then render the final version at higher quality.
- Write the prompt like a director: subject, setting, camera move, mood, and any dialogue.
- Use image-to-video when you want a specific look: give it a start frame (and optionally an end frame).
- Download the clip and finish it in an editor like CapCut.
Try this
A barista slides a latte across a wooden counter in a sunlit cafe, slow push-in, warm morning light, soft ambient cafe noise, she smiles and says 'Have a great day.'
Prompt pattern: subject + setting + camera + light + sound.
Why it matters for business
Insurance and business angle: short branded explainer clips, social ads, and product demos. Always label AI video where required and follow platform rules.
Full walkthrough: from idea to a finished clip
Video models are expensive to experiment with if you go in unplanned. This process keeps costs low and results predictable. Check bfl.ai for the current buttons and prices, since interfaces change.
1. Decide what the clip must do
Write one sentence: "This 10-second clip makes a viewer feel ___ and then do ___." Example: "This clip makes a viewer hungry and then visit the bakery." If you cannot write the sentence, do not generate yet.
2. Plan shots before you prompt
Break the video into 1 to 3 shots. For a 10-second ad: shot 1 is the product close-up, shot 2 is a person enjoying it, shot 3 is the logo. Shorter shots are easier to get right than one long scene.
3. Write the prompt like a director
Use five parts in this order. Each part removes guessing for the model:
- Subject: who or what, and what they are doing.
- Setting: where, and what time of day.
- Camera: slow push-in, handheld, wide shot, close-up, orbit.
- Light and mood: warm morning light, moody neon, bright and clean.
- Sound: ambient noise, music feel, and any line of dialogue in quotes.
A baker pulls a tray of croissants from a stone oven in a small bakery at sunrise. Slow push-in on the croissants. Warm golden light, steam rising. Soft oven crackle and quiet morning street noise. The baker smiles and says, 'Fresh today.'
4. Test in Draft mode first
The page notes a cheaper Draft mode. Use it to test the idea. Generate two or three versions with small changes in the prompt, change only one thing at a time (camera, then light, then wording) so you learn what each change does. Keep a note of what worked.
5. Use a starting image when the look matters
Image-to-video begins with a picture you provide, such as your real product photo or a logo scene. This keeps the brand look consistent. Describe only the motion in the prompt, because the picture already says what things look like.
6. Render the final, then check it
Once a draft looks right, render at higher quality. Then watch the whole clip twice, once with sound on and once muted. Use this checklist:
- Hands, faces and text look normal and do not change shape.
- No fake logos, real people, or brand names you do not have rights to.
- Nothing in the clip makes a claim you cannot prove (prices, results, guarantees).
- The audio matches the picture.
7. Finish and publish
Download the clip, then add your logo, captions and call to action in CapCut. Label it as AI-generated where the platform or law requires it.
Common problems
- Weird or melting details: simplify the scene, reduce the number of people, shorten the clip.
- Text in the video is garbled: add text later in CapCut instead of asking the model to draw it.
- Result ignores your camera move: use one clear move per shot.
- Costs add up: set a budget before you start, such as "ten drafts and two finals."
Business use case
Scenario: A local business owner (a bakery, a gym, a real estate agent) needs a short video for an upcoming promotion but has no camera crew or budget for a shoot.
How it works: Write a prompt describing the product, the setting and the mood, and ask for a short clip with sound. Generate a few versions, pick the best, then add the logo, offer and call to action in an editor like CapCut.
Measure success by: Cost and time per finished video compared with a traditional shoot, and views, clicks or inquiries from the post.
Watch out for: Do not show real people, trademarks or claims you cannot support. Check every clip for odd details before you post, and follow the platform's rules on labeling AI-made content.
Quick facts
- Announced July 23, 2026 in early access; FLUX 3 Video opened publicly in early August 2026.
- Up to 20 seconds per generation, 720p to 1080p, with audio.
- Modes: text-to-video, image-to-video (start/end frames), video continuation.
- Reported launch pricing: about $0.06 per second for Draft, higher for HD and Full HD. Confirm current prices on bfl.ai.
- An open-weight FLUX 3 Dev version has been announced for later.
What to learn next
Try the same prompt in FLUX 3 and Higgsfield and compare. Then learn what an API is (see the glossary) so you can call FLUX from code.