FLUX.3 Video Prompt Guide: Shot Briefs, Timed Beats, Sound and Real Examples

All guides · FLUX.3 Video in Fuser

Quick answer: The best way to prompt FLUX.3 Video in Fuser is to write it like a shot brief: camera, subject and action, setting and light, then the sound, with timed beats once a clip runs past a few seconds; draft the idea cheaply, then render the keeper on FLUX.3 Max at 1080p.

FLUX.3 Video is Black Forest Labs’ video model. In Fuser, one FLUX.3 Video node covers text-to-video, a start frame, a start and end frame and extending a clip, from 5 to 20 seconds at 720p or 1080p, with sound made in the same pass; it picks the route from what you connect. Black Forest Labs’ text-to-video guide treats a prompt as directing a scene rather than listing objects. Every example below is a real prompt we ran on 23 October 2026, shown with its output.

The FLUX.3 Video prompt formula

Black Forest Labs’ video prompting overview lists what a prompt should cover. Write it in roughly this order:

1. Camera. Shot size, angle and movement: “handheld medium-wide shot”, “a slow push-in”, “locked-off”.

2. Subject and action. Who moves and exactly what happens, in the order it happens.

3. Scene and light. Place, time of day, weather, and where the light comes from.

4. Motion and style. How things move (slow, abrupt, weightless) and the look: film stock, palette, grain.

5. Sound. The sources you want to hear, and any spoken line in quotation marks.

With a start frame or a clip to extend, add what must stay the same. Black Forest Labs’ suggested default is one line in a fixed shape, so you can revise one part without touching the rest: “[camera] shot of [subject] [action] in [environment]. [supporting visual and motion details]”. Short prompts leave framing, motion and mood to the model; long ones pin them down, though the guide warns that over-stuffing can make motion less coherent. The dancer prompt under Draft below follows that shape.

Timed beats for long takes

Set Duration from 5 to 20 seconds (the default is 5). When action has to land on time, Black Forest Labs recommends timestep prompting: two or three beats for a 5-second clip, and a labelled hard cut wherever the angle changes. For one unbroken take, give each window its own action and put the camera move inside its beat.

Timed beats (text-to-video, 15 s, 16:9, 1080p, FLUX.3 Max): “One unbroken handheld take on 16mm film, misty grey dawn, a muted palette of wet timber, slate water and fog with one faded-red accent. 0-4s: inside an old timber boathouse on a still lake, a young woman in a faded red rowing vest and grey leggings lifts a pair of long wooden oars down from a wall rack. 4-8s: she carries the oars on her shoulder down the wet wooden slipway, the camera following behind her at shoulder height. 8-12s: she steps carefully into a narrow white single scull tied at the jetty, sits, and slots the oars into their riggers. 12-15s: she pushes off from the jetty with one hand and takes her first stroke, the scull gliding out into the fog, ripples spreading toward the camera. Sound: creaking boards, her footsteps, water lapping, one distant bird call.”

One take with no cuts, and the first beats landed on time: the oars came off the rack in the first four seconds, and by eight she was out at the scull with the camera behind her. The look is what the first sentence asked for: slate water, fog, wet timber and one faded-red vest. The props slipped later. The scull was already rigged, so one oar was left standing on the jetty, and she sat facing the bow. When the logic of objects matters, spell out their state and the body position, for example “an empty single scull with bare riggers” and “she sits facing the stern”, and check the take.

Start frame: describe the motion, not the picture

Connect a Start Image and the node runs image-to-video; with Aspect Ratio on Auto the clip takes the image’s shape. Black Forest Labs’ image-to-video guide calls this the most literal option: the image becomes the opening frame and the prompt drives the motion. The picture is decided, so spend the words on what moves, the camera and the sound.

Start frame (image-to-video, a bookshop portrait as the start image, 8 s, 1080p, FLUX.3 Max): “He takes off his round tortoiseshell glasses and wipes the lenses on his rust sweater, then turns away to the shelf on the right and pulls out a faded green book. He turns back to the camera, puts the glasses back on, opens the book and looks up with a broad smile. The camera holds a steady medium shot with a slight drift. Sound: quiet shop room tone, the book sliding from the shelf, pages turning.”

He stayed the same man for all eight seconds: beard, cap, rust sweater, tortoiseshell frames, even the mole on his cheek. He wiped the lenses on his sweater and kept the glasses in hand while he reached for the shelf. The book came out dark rather than faded green, and he opened it a beat before putting the glasses back on. The first frame was a close redraw of our photo, with a touch more contrast, rather than the file itself, so grade the clip if it must cut against the original still.

Start and end frames

Add an End Image and the node bridges the two; the end image needs a start image. Black Forest Labs advises keeping both frames related, with the same subject, scene or camera set-up: the bigger the gap, the more freedom the model takes and the less predictable the result. Prompt the change between the frames, not the frames.

Start to end (start = a peony bud, end = the open bloom, 5 s, 1080p, FLUX.3 Max): “Locked-off camera with a very slow push-in. Time-lapse over one day: the coral peony bud loosens and opens petal by petal into a full, heavy bloom, its head dipping on the stem; two petals drop and land on the oak table; the sunlight and the shadow on the plaster wall slide slowly to the right. Sound: quiet room tone, a soft tap as the petals land.”

It opened the bud petal by petal, let two petals fall and settle where our end frame has them, and finished on the end composition, with the shadow growing on the wall. As with the start frame, both ends came back slightly tighter and punchier than our files. The base FLUX.3 model takes a start and end pair too.

Dialogue and sound

Generate Audio is on by default, and the sound is made with the picture. Black Forest Labs’ audio and speech guide gives the rules that matter:

• Quote the exact words and say who speaks them; call an off-screen line a voiceover.

• Use up to four layers, speech, ambience, effects and music, but only the one or two the scene needs.

• Name sources, not moods: “distant traffic and rain ticking against a metal awning”, not “moody ambience”. “Quiet room tone” may come back as static or dead air.

• Timing is a target, not a control. If the last word is cut off, shorten the line or lengthen the clip.

Camera move and a line (text-to-video, 5 s, 16:9, 720p, FLUX.3 Max): “Single continuous camera move, no cuts. The shot opens wide on a young florist with curly black hair and a mustard canvas apron, standing among galvanised buckets of coral ranunculus and white anemones under the iron-and-glass roof of an empty covered market at first light. The camera arcs a slow half circle around him, from behind his left shoulder round to face him, while pushing in, and ends in a close-up as he looks up from the stems in his hands and says: "First ones in today." Cool blue daylight from the high windows, one warm bulb above the stall. Sound: the rustle of paper and stems, a distant shutter rolling up, his line spoken clearly.”

The camera did exactly what we wrote: from behind his shoulder, round past the coral ranunculus and white anemones, into a close-up as he looks up. The line did not fit. His speech starts at about 4.4 seconds of 5, and our transcript caught only “in today” before the clip ended. Give a line room: put it early with a time (“At 2s he looks up and says…”), cut the action before it, or set the clip to 7 or 8 seconds. Our FLUX.3 Video tracks also came back mono and mixed low, so plan to level them in the edit.

Text on screen

Black Forest Labs lists on-screen text among FLUX.3 Video’s abilities: give the exact words in quotation marks and say where they sit. The reverse also applies: a quoted line with no speaker may be drawn as a subtitle, so add “no on-screen text or subtitles” when you want speech only.

Text test (text-to-video, 5 s, 16:9, 720p, FLUX.3 Max): “Slow sideways dolly along a pale mint-green ferry ticket kiosk on a stone quay at hard noon, the sea bright behind it. Painted on the kiosk in large white capitals: "HARBOUR FERRY". A small black chalkboard on the counter reads "NEXT BOAT 4:15" in white chalk. A ticket seller in a navy cardigan slides a paper ticket across the counter. 2-4 s: a subtitle in small white sans-serif, centred at the bottom of the frame: "Last crossing of the summer." Sound: gulls, water slapping the quay. No background music.”

“HARBOUR FERRY” and “NEXT BOAT 4:15” came out spelled right and held through the dolly, and the subtitle was spelled right in every frame we checked. Its timing was not: we asked for 2 to 4 seconds and it ran from about 0.75 seconds to the end. When words must come and go on cue, add them as a text layer in Fuser’s Compositor instead.

Draft first, then the final

With Model set to FLUX.3, switch Draft on for a fast 720p take; resolution is fixed at 720p in Draft. A Draft second costs about a fifth of a 1080p second. When the idea works, switch Draft off, or set Model to FLUX.3 Max, which has no Draft, and render at 1080p.

Draft, then final (text-to-video, 6 s, 16:9; Draft on FLUX.3, then FLUX.3 Max at 1080p): “Handheld medium-wide shot of a young dancer in a faded navy T-shirt and charcoal sweatpants rehearsing alone in a sunlit wooden studio: she spins twice across the worn floorboards, one untied shoelace whipping out behind her, then drops low into a slide and rises with her arms opening toward the tall windows. 35mm colour-negative look, soft window light from the left, slight motion blur on her hands, a ballet barre and a wall of mirrors behind her. Sound: sneakers squeaking on the boards, her breath, a radiator ticking.”

Both takes followed the brief: two spins, the drop into a low slide, the rise with her arms opening to the windows, sneakers squeaking on the boards. They are not the same shot, though. The final framed the room differently, put patches of sun on the floor and ended facing the camera. Fuser’s node doesn’t carry a draft over into the final render, so use drafts to settle the action, camera and timing, not one particular frame.

Extend a clip

With Model set to FLUX.3, connect a clip to the Video input and the node continues it from its last frames, following your prompt. It can’t be combined with start or end images, and the clip must be an MP4 under 15 seconds and 50 MB. Black Forest Labs says the source sets the motion, subjects and look; the prompt only steers how the shot carries on, so restate the place, the light and the person.

Extend (source = the 6 s FLUX.3 Max dancer take above, 5 s, 1080p): “The same dancer in the same sunlit wooden studio lets her arms fall, catches her breath, then walks to the tall windows and picks up a water bottle from the sill and drinks, the handheld camera drifting after her at the same height. Same soft window light from the left and 35mm colour-negative look. Sound: her footsteps on the boards, her breathing slowing, the radiator ticking.”

The extension picked up on her last pose with no visible jump, walked her to the window and had her drink from a bottle on the sill, the camera following. Its colour ran a touch warmer than the source. The node returns only the new five seconds; we joined them to the source in the video above. Extend is FLUX.3’s priciest mode, close to twice its normal 1080p rate per second, so try the next beat with Draft on first.

Fix what goes wrong

• Fast physical action: our judo throw stalled halfway and restarted, and the thrown woman landed on her side instead of her back. Keep to one move per clip and describe the body mechanics in order. For fast contact between two people, Kling 4.0 did better on the same prompt in our FLUX.3 Video vs Kling 4.0 test.

• Props that don’t add up: the rower’s scull came pre-rigged, so a carried oar had nowhere to go. State what each object looks like before the action starts.

• Late or clipped lines: a 5-second clip left our line no room. Time the line early, keep it short, or lengthen the clip.

• Text that stays too long: the subtitle ran most of the clip, not 2 to 4 seconds. Overlay timed text in the Compositor.

• Extra people in reflections: in both dancer takes, the studio mirror showed figures who weren’t in the room. Say what the mirror shows, or frame away from it.

• Redrawn frames: start and end images come back as close redraws with a little more contrast, not the files themselves.

• Quiet sound: our tracks were mono and mixed low. Name concrete sources, and level the clip in the edit.

Fast action (text-to-video, 5 s, 16:9, 1080p, FLUX.3 Max): “Real-time speed, 35mm documentary look, flat daylight from tall paper windows in an old wooden dojo with straw-yellow tatami. Two women judoka, one in a white judogi and one in a deep blue judogi, grip each other's lapels and sleeves. The woman in white steps in, turns her hip under her opponent and throws her in a clean hip throw; the woman in blue rotates over her hip through the air, lands flat on her back with a loud slap of the mat and immediately rolls up to her knees. Camera: low handheld, side-on, following the throw. Sound: bare feet on tatami, a sharp exhale, the slap of the landing.”

The dojo is exactly as written, paper windows and straw-yellow tatami, and the slap of the landing is the loudest sound in the clip. The throw itself is the weak point: watch the middle second.

FLUX.3 Video settings in Fuser

The FLUX.3 Video node exposes:

• Model: FLUX.3 or FLUX.3 Max. Max reasons about your prompt first; FLUX.3 adds Draft and Extend.

• Prompt.

• Start Image and End Image: image-to-video; the end image needs a start image.

• Video: a clip to extend, FLUX.3 only, not combined with images.

• Draft: a fast, fixed-720p take, FLUX.3 only.

• Duration: 5 to 20 seconds, default 5.

• Aspect Ratio: Auto, 21:9, 2:1, 16:9, 4:3, 1:1, 3:4 or 9:16.

• Resolution: 720p (default) or 1080p. At 16:9 our clips came back 1280×704 and 1920×1088; Black Forest Labs’ FLUX.3 Video docs say output sizes are multiples of 32, so crop if you need an exact frame size.

• Generate Audio: on by default.

• Reasoning Level and Seed: FLUX.3 Max only; Reasoning Level is Low or High (the default).

FLUX.3 and FLUX.3 Max cost the same per second, and the node shows the cost before you run it (Fuser docs).

Build it as a workflow

FLUX.3 Video does its most predictable work from frames you control. Make the start and end frames with an image model such as FLUX.3 Image (see the FLUX.3 Image prompt guide), wire them into FLUX.3 Video, and keep the prompt in its own Text node so a Draft node and a FLUX.3 Max node can share it. Feed the keeper back into the Video input to extend it. For shots that depend on references or on clips longer than 20 seconds, the Kling 4.0 prompt guide covers the same ground for Kling 4.0.