Creators test Kling 4.0 with dialogue and tracking-shot prompts

• Creators are scripting complete scenes: MayorKingAI's kitchen prompt specifies eight shots, dialogue, reaction beats and lighting continuity.
• Phone-video realism is being scripted too: minchoi's brunch example requests a bumped phone, shifting auto-exposure and café microphone noise.
• Action tests use different inputs: the ski example starts with a character sheet, while the skateboard example uses text alone for a continuous tracking shot.

One kitchen-drama prompt fixes the hallway light at 5600K and assigns each actor a permanent side of the screen. A concert test schedules mouth-closed instrumental gaps, while Kling's official guide warns that music and subtitles can appear by default.

MayorKingAI published a 30-second scene about a husband confessing that he lost his job, praising its dialogue, lip sync and facial emotion.

The creator specifies eight shots in the accompanying prompt:

1. Wide: Daniel grips a glass as Lucía enters the kitchen.
2. Slow push-in: he admits he was fired in June.
3. Reverse: she freezes, grips a chair and confronts him.
4. High-angle close-up: he explains his months of pretending to commute.
5. Silent reverse: her grip loosens as anger becomes hurt.
6. Medium: she sits and asks why he did not trust her.
7. Silent close-up: he tries to answer but cannot.
8. Profile two-shot: she promises support and takes his hand.

The silent reaction shots are the smartest part of the prompt. Its acting directions specify swallowing, irregular blinks and a trembling chin, with clock ticks and refrigerator hum underneath the dialogue.

One reproducibility snag: the final constraint requests four dialogue turns, but the shot list scripts five, two for Daniel and three for Lucía.

Minchoi shared four prompts featuring a character named Sadie in 30-second, vertical, continuous takes. All request soft phone imagery, compression, no color grading and location-specific microphone sound.

• Brunch: a waiter bumps the table around nine seconds; an off-camera sister steals a pancake bite around 26 seconds.
• Airport: Sadie sips coffee, reacts to a boarding announcement and yawns mid-sentence in the airport prompt.
• Parked car: rain drums on the roof, her voice sounds close and boxy, and a passing trolley prompts her to duck out of sight in the car prompt.
• Beach: auto-exposure changes as her head blocks the sun, while gusts partly swallow her voice in the beach prompt.

Minchoi acknowledged in a reply that some subtle details still felt off.

The ski example starts from a character sheet and requests the same outfit and gear throughout. Its prompt fixes the sun on the skier's left and divides the 30-second descent into seven viewpoints:

1. 0–4 seconds: drone follows the drop from the ridge.
2. 4–8: helmet camera follows the turns.
3. 8–12: pole-mounted camera points back at the skier.
4. 12–17: a camera planted in the snow catches an overhead jump.
5. 17–21: a chase skier follows from behind.
6. 21–25: a distant telephoto camera catches a backlit carve.
7. 25–30: a low static camera receives a blast of snow from the final stop.

Audio directions distinguish wind hitting the microphones, ski-edge chatter, landing impacts and breathing, with music kept underneath.

Minchoi describes the skateboard test as text-only. The prompt divides one continuous low tracking shot into five six-second blocks:

1. 0–6 seconds: ride toward the camera, push once and center weight.
2. 6–12: approach stairs, bend knees and lift the board's nose.
3. 12–18: clear the stairs and absorb the landing.
4. 18–24: roll out, check speed and carve left.
5. 24–30: widen as the skater rides away.

It explicitly requests consistent clothing and facial identity, cloth and hair responding to motion, and natural board sound.

Minchoi labels the concert example “Kling 4.0 + Suno v6.” The two published prompt blocks run 28.7 and 12.4 seconds, with separate instructions for vocal timing and broadcast-camera cuts.

• Tempo: 98 BPM, with each bar specified as 2.45 seconds.
• Vocal timing: the first eight seconds are instrumental; subsequent lyric lines receive individual time windows.
• Continuity: the microphone stays in the singer's right hand, and the band remains hidden upstage.
• Broadcast texture: exposure shifts, focus corrections, long zoom lenses and handheld shake are requested alongside continuously moving crowds.

Magnific's contribution is a 12-second miniature chase.

An early-access test from techhalla takes the form of a Cadillacs & Dinosaurs homage. The creator later described Kling as versatile for action and prompt adherence, but said in a reply that it required changing how he prompted.

techhalla's Cadillacs & Dinosaurs homage

Kuaishou dates Kling 4.0's launch to October 15 in its official announcement. The user guide lists:

• Duration: 3–30 seconds per generation, adjustable in one-second increments.
• Prompt length: up to 8,000 tokens.
• References: up to 15 combined assets, including images, videos, voice profiles and elements.
• Keyframe control: up to 10 keyframe inputs.

The guide marks 10-bit HDR output and extensions up to two minutes as coming soon; Kuaishou also describes API access as forthcoming.

For cost, CinemaDrop's October 16 guide, citing a launch-day interface check, reports 450 credits for a 30-second 1080p generation or 300 credits at 720p.