What FLUX 3 Video Can Do: 8 Real Clips with Prompts and Results

👁 119 reads

What can FLUX 3 Video really do, Black Forest Labs’ video model sitting at 6th place on the Artificial Analysis leaderboard? I generated 8 clips across varied use cases - from spoken foreign-language dialogue to chalkboard lettering and animating a still photograph. Below you will find the complete prompts, parameters, costs, and an honest breakdown of what worked and what fell flat. At the end, I have also included six official clips from the developer demonstrating features I did not test directly. Full pricing, rankings, and an setup guide are available in the FLUX 3 Video guide.

1. Local business commercial with spoken dialogue

Use case: A short social media ad for a bakery, cafe, or repair shop - featuring a subject speaking directly to the camera.

Prompt: (Note: The prompt was formulated in English, with the dialogue requested specifically in Polish as a foreign-language test):

Short TV commercial for a Polish bakery. Medium shot of a smiling female baker in her fifties in a small sunlit bakery, holding a fresh loaf of rye bread toward the camera. She says clearly in Polish (translated from the Polish original): "Sourdough bread, baked today at five in the morning." Behind her, shelves full of bread and a window with a quiet street. Sound: her voice, soft crackle of the crust, quiet street outside, no music. Realistic, warm light, no subtitles, no on-screen text.

Verdict: Whisper transcribed the line word-for-word as (translated from the Polish original) “Sourdough bread, baked today at 5 a.m.” (“Sourdough bread, baked today at 5 a.m.”) - exactly as requested, with the digit inserted by Whisper itself. The baker raises the loaf toward the camera, shelves are packed with bread, and a quiet street sits outside the window.

2. Speech rendering: Language tongue-twister

Use case: A test to see whether the model can handle complex foreign-language phonetics without garbling.

Prompt: (Scene 2 from our benchmark arena; the prompt specifies a notoriously difficult Polish phonetic tongue-twister):

Medium close-up of a woman in her thirties sitting at a kitchen table, looking straight into the camera. She says clearly and calmly in Polish (translated from the Polish original): "In Szczebrzeszyn a beetle buzzes in the reeds, and I pay thirty-two original Polish currency units." Natural daylight from a window, static camera, realistic, no subtitles, no on-screen text, no music.

Verdict: Whisper transcription (translated from the Polish original): “In Szczebrzeszyn a beetle buzzes in the reeds, and I pay 32 original Polish currency units.” - zero word errors. For comparison, Seedance 2.5 mangled nearly every second word in this same scenario (a WER of 0.583), while Wan 3.0 scored 7.5/10. The entire sentence fits neatly into the 5-second duration.

3. On-screen text: Classroom chalkboard

Use case: Shop signs, boards, brand names - anything that requires crisp, legible typography in-frame.

Close-up of a green school chalkboard in a sunlit classroom. A hand finishes writing with white chalk. The chalkboard clearly shows exactly two lines of text in capital letters: first line (translated from the Polish original) "VIDEO ARENA 2026", second line (translated from the Polish original) "YELLOW BOAT". Static camera, realistic, chalk dust in the light, no other text.

Verdict: Flawless. The hand wraps up writing the final letter (translated from the Polish original letter), and both lines match the prompt precisely - complete with complex accents and diacritics. In our arena tests, only Wan 3.0 and MiniMax H3 Max managed these characters correctly previously. Seedance 2.5 hallucinated [corrupted lettering in the original Polish test], while Kling 3.0 and Veo 3.1 Lite fared worse.

4. Camera movement: Motorcycle orbit

Use case: 360-degree product showcases, automotive shots, and furniture reels.

A smooth 180-degree camera orbit around a vintage red motorcycle parked on a wet cobblestone street in an old European town at blue hour. The camera circles steadily from the front of the motorcycle to its back at a constant speed. Street lamps reflect in the puddles, light drizzle. The motorcycle keeps its exact shape and details. Realistic cinematic footage, real-time motion, no text, no logos.

Verdict: The camera executes a full 180-degree sweep from front to rear at an even pace, while the bike maintains its shape, chrome detailing, and headlight form without warping. Street lamps reflect cleanly in the damp cobblestones. One flaw: the requested light drizzle is barely visible. Overall, 8.5/10 - matching Wan 3.0 and Kling 3.0 in arena testing.

5. Fluid physics: Water poured until overflowing

Use case: Beverage commercials, kitchen sequences, or any dynamic fluid interaction.

Static close-up on a wooden kitchen table in daylight. Water is poured from a glass pitcher into an empty clear drinking glass. The stream splashes, bubbles rise, the water level rises steadily until the glass overflows and water spills onto the table and spreads into a puddle. Real-time speed, no slow motion, realistic physics, no text.

Verdict: The entire chain of events unfolds in order: empty glass, bubbling liquid, climbing water level, overflow, and a puddle forming on the surface. In our previous benchmarks, only Wan 3.0 managed this sequence. Toward the end, the stream rolls down the outside of the glass with a slight gel-like consistency - hence an 8.5 rather than a 9.

6. Packshot with sound: Espresso machine

Use case: Product commercials requiring rich foley audio (ASMR, appliance mechanics).

Cinematic product commercial. Close-up of a matte black espresso machine on a marble kitchen counter at sunrise. Hot espresso streams into a small white cup, steam rises through a beam of golden light, the camera slowly pushes in. Sound: the hiss of the machine, liquid pouring, cup clinking against saucer, soft ambient morning birds. Realistic, shallow depth of field, no text, no logos.

Verdict: Pure commercial aesthetic: polished marble, sunrise glow, rising steam, rich crema, and a slow push-in. The audio layer, however, was only partially delivered. The track contains birdsong and background ticking, but completely omitted the machine’s hiss and pouring sounds. The audio track is also quiet (averaging -38.6 dB). For commercial projects, you will want to layer foley in post-production.

7. Image-to-video: Cafe interior

Use case: Architectural visualizations, interior showcases, or turning still product photos into video.

Starting image: An interior visual generated during my earlier FLUX 3 Image interior tests.

The camera slowly dollies forward through the cafe toward the curved terracotta counter. The barista behind the counter grinds coffee, a customer walks past the counter, the woman at the left table lifts her cup and laughs with her friend. The pendant lamps hang still. Sound: cafe ambience, a coffee grinder, soft chatter, cups clinking. Realistic, natural motion, no text.

Verdict: A smooth dolly move toward the counter: a customer walks across frame, a barista appears behind the bar grinding coffee, and the scene preserves the tiling, light fixtures, and plants from the reference still. Note that with the aspect ratio set to “auto”, the model cropped the frame to 1088×800, trimming portions of the ceiling and sides; specify the aspect ratio manually if you need the entire composition preserved. I fed the exact same prompt and image into Seedance 2.5 in our head-to-head comparison.

8. Consistent character across two shots

Use case: Short narrative films, character-led ads, and episodic AI productions.

Two shots separated by one hard cut. Shot 1: wide shot of a young woman with short red hair, round black glasses, a yellow raincoat and a black backpack walking along a rainy city street at night with neon reflections on the wet pavement. Hard cut. Shot 2: close-up of the same woman with the same short red hair, the same round black glasses and the same yellow raincoat as she stops under a shop awning and looks up at the rain. Realistic, cinematic, no text.

Verdict: The hard cut lands right on cue. In the close-up, the character displays the same red hair, circular black frames, yellow raincoat, and backpack. In the wide shot, she is walking away from the camera, leaving facial details concealed - an expedient framing choice by the model. Reflections across wet asphalt look convincing.

Official materials: Features I did not test

The following clips were shared in Black Forest Labs’ post from August 4, 2026, and are streamed directly from their media servers. These are producer demos; BFL did not release the underlying prompts.

Keyframe transitions: Empty room to furnished living space

First-and-last-frame interpolation is particularly valuable for interior designers: feed it a “before” and “after” image, and the model synthesizes the camera path in between. On fal.ai, this costs the same as standard generation: $0.085/s.

Video extension: Moose on the road

The model ingests up to 4 seconds of existing footage alongside audio and continues the shot while preserving camera trajectory. This is the pricier mode: $0.41/s at 720p on BFL’s platform.

Multi-shot continuous generation

Period-accurate dialogue

World knowledge: Freiburg in few words

BFL notes that the model pairs training weights with “real-time grounding,” designed to assist in factual or documentary video creation.

Draft mode versus full production quality

Draft mode maintains overall composition and motion paths, letting you preview several concepts before spending full render credits. On fal.ai, draft mode runs at $0.03/s.

Key takeaways

• Strengths: Multilingual speech delivery and on-screen typography (including tricky diacritics), realistic liquid physics, and character continuity across cuts.
• Weaknesses: Foley sound precision (the coffee maker sequence missed the machine noise), and subtle prompt cues (mizzling rain, auto-crop quirks).
• Pricing: $0.85 per 5 s on OpenRouter, or $0.425 via fal.ai. Pricier than Wan 3.0 at 480p, but rendered natively in 720p.

For side-by-side analysis, check our FLUX 3 Video vs. Wan 3.0 vs. Seedance 2.5 benchmark, or inspect the script used here in the FLUX 3 API tutorial.