The Best Way To Prompt Flux 3 For AI Ads [Example + Full Prompts Inside]

We got Flux 3 early at Maxfusion AI, ran thousands of generations through our Agent, and kept testing inside the platform. Here is the prompting method that wins for text-to-video ads and image-to-video.

It’s the timestamp method, but stripped down. Two parts, one hard ceiling, and one thing Flux 3 does better than any other model.

What Flux 3 actually is

First, the positioning, because it changes how you prompt.

Seedance 2.5 is the strong reasoning model. It handles cinematic work, complex staging, long causal chains. You can write it a dense prompt and it will follow the logic. To read how to best prompt it, see the full Seedance 2.5 breakdown.

Flux 3 is different. Its secret weapon is era mimicry. It reproduces the ‘90s and 2000s better than anything we’ve tested at Maxfusion AI: standard-definition broadcast video, daytime TV lighting, VHS-era texture, the whole period look.

Prompt it accordingly. Super-simple beats, period framing, minimal action. That’s where Flux 3 delivers its most consistent results, generation after generation.

Part 1: the opening block

Three or four lines before any timestamps. No more.

The opening block covers exactly four things: the type of video it is, the era and how it was shot, the set and the elements in it, and the characters.

This block does the nostalgia work. “Early-2000s daytime courtroom television show, shot on standard-definition broadcast video. Bright flat lighting, warm oversaturated wood tones, mild grain.”

Era first, then set, then people. Once Flux 3 locks onto the period, the lighting, texture, and camera behavior follow on their own.

Part 2: the timestamps

After the opening block, the video is written as simple timestamped blocks.

Inside each block, three things: what the camera shows, what happens, and the exact dialogue with who says it.

Dialogue is word for word and always attributed. “The judge lowering her glasses: ‘Ma’am, it’s a soda.’” Never a summary like “the judge dismisses her.” Flux 3 renders the line you wrote, so write the actual line.

Keep the blocks simple. One camera setup, one event, one exchange per block.

The ceiling: 3 actions

Flux 3 has a hard ceiling, and respecting it is most of the method.

Two to three actions total per generation. Not per block, per video. If you write more, it struggles: actions blur together, timing slips, consistency drops.

This is the biggest difference from prompting Seedance 2.5, where you can chain complex action. On Flux 3, when in doubt, cut an action. The simple version generates clean, and the packed version does not.

The full example

This prompt follows the method exactly: opening block, timestamped events, attributed dialogue, three actions.

Early-2000s daytime courtroom television show, shot on standard-definition broadcast video. Bright flat lighting, warm oversaturated wood tones, mild grain. A small TV courtroom set with a raised judge’s bench, an American flag, and two podiums facing the bench. Locked-off multi-camera coverage, no handheld shake. A stern female judge in her 60s with reading glasses and black robes. At the plaintiff podium, a woman in her 20s in a cardigan, arms crossed. At the defendant podium, her roommate, a woman in her 20s in a hoodie, avoiding eye contact. A bailiff stands to the side. Scene starts mid-case.

0-6s. Medium shot on the plaintiff, jabbing a finger toward the defendant: “Your honor, it was the LAST strawberry lemon. I had my name on it. On tape. There was tape, your honor.”

6-12s. Cut to the judge lowering her glasses: “Ma’am, it’s a soda.” The plaintiff, immediately: “It is NOT a soda, your honor. It’s a prebiotic.”

12-18s. The bailiff hands the judge a bright pink-and-yellow can. Close-up: the can reads “poppi” in lowercase letters with “strawberry lemon” beneath it. The judge turns it, reading the back, eyebrows rising: “Five grams of sugar… this is good for your gut?” She looks at the defendant over her glasses.

18-20s. Wide shot. The judge sets the can down and bangs the gavel once: “Buy her a twelve-pack. Judgment for the plaintiff.” Courtroom murmur, gavel echo, no music.

The do and don’t list

• Do open with 3 to 4 lines: video type, era and how it was shot, set, characters.

• Do lean into the ‘90s and 2000s. Period formats are where Flux 3 outperforms everything else.

• Do write every line of dialogue word for word, attributed to the speaker.

• Do state what the camera shows at the start of each timestamp block.

• Do count your actions and stop at 3.

• Don’t write cinematic, complex-staging prompts. That’s Seedance 2.5 territory.

• Don’t summarize dialogue. Flux 3 renders exactly what you write.

• Don’t stack events inside one timestamp block.

• Don’t pad the opening block past 4 lines. The nostalgia work needs density, not length.