Kling v3 by Kuaishou is a text-to-video model built for photorealistic scenes. It is available in GPTunneL: you write a prompt, pick the resolution (720p or 1080p) and the clip duration, and get a finished video. This guide covers the part that matters most — how to phrase the prompt so the model shoots exactly the scene you had in mind.
Anatomy of a video prompt
A good Kling v3 prompt reads like a brief for a camera operator: what is in the frame, what happens, and how it is filmed. Build it from five layers.
1. Scene and subject
Start with who or what is in the frame and where. One main subject, a concrete location, a couple of clarifying details: "an elderly potter in a workshop, hands covered in clay," not just "a person working."
2. Action and motion
Video is about change over time. Describe one continuous action: "the potter slowly pulls up the walls of a vase as the wheel spins." Without an action, the model tends to produce a nearly static "living photograph."
3. Camera: shot size, movement, lens
Specify the shot size (close-up, medium, wide), the camera movement (static, slow push-in, left-to-right pan, orbit), and the character of the lens (shallow depth of field, wide angle). This is the most underrated layer — it is what makes a clip look filmed rather than generated.
4. Light and atmosphere
Kling v3's photorealism shines through lighting: "soft morning light from a window," "neon reflections on wet asphalt," "golden hour." Add weather and the mood of the scene.
5. Style
Close the prompt with a style frame: "cinematic, photorealistic," "documentary look," "commercial product shot." One or two phrases, not a list of ten adjectives.
How a video prompt differs from an image prompt
- One continuous action, not a list of objects. An image prompt is an inventory of what should appear in the frame. A video prompt is a micro-script: the subject does one thing from the start of the clip to the end.
- Timing. Think about what happens over the course of the clip: "the candle burns down," "a wave rolls in and pulls back." Words like "slowly," "smoothly," and "gradually" control the pace.
- Camera movement is part of the scene. In an image the angle is fixed; in video the camera lives — push-in, pull-back, pan. If you don't specify it, the model decides for you, and not always well.
Techniques
- Verbs over adjectives. "Steam rises from the cup as a woman takes a sip" works better than "a beautiful cozy coffee scene." Motion is carried by verbs.
- One narrative beat per clip. One scene, one event. Don't ask for "he walks in, sits down, opens a laptop and makes a call" — you'll get a mess. Split it into separate generations.
- A series of consistent shots for editing. Need something longer? Generate several clips that share a "scene passport" — repeat the description of the character, location, light, and style in every prompt, changing only the action and the shot size. Then cut them together in an editor.
- Match resolution and duration to the job. Use 720p for fast iterations while you tune the prompt, and 1080p for the final render. A short duration holds a single action better; go longer only when the motion genuinely develops over time.
Ready-to-use prompts — copy and adapt
Product clip:
A perfume bottle on a glossy black surface, thin mist slowly swirling around it. The camera orbits the bottle in a smooth arc, close-up, shallow depth of field. Cool studio lighting with a blue rim light. Commercial product shot, photorealistic.
Landscape with camera movement:
A mountain valley at dawn, fog drifting low over the ground. The camera flies slowly forward over a river at low altitude. Golden hour, long soft shadows, light haze. Cinematic, photorealistic.
Character scene:
An elderly fisherman in a yellow raincoat stands at the bow of a boat, slowly hauling a net out of the water. Medium shot, static camera, the boat rocking gently. Overcast morning, diffused light, light drizzle. Documentary look, photorealistic.
Atmospheric b-roll:
A city at night after rain, neon signs reflected in puddles. The camera pans slowly from left to right at ground level. Pedestrians out of focus in the background. Cinematic b-roll, photorealistic.
Macro detail:
A drop of water slides slowly down a monstera leaf and falls. Extreme close-up, static camera, shallow depth of field. Soft diffused daylight. Photorealistic macro shot.
Common mistakes and how to fix them
- An overloaded scene. Three characters, five objects, and two events in one prompt — the model spreads its attention and fails at everything. The fix is subtraction: one subject, one action, everything else stays background.
- Contradictory motion. "A static camera slowly orbits the object" or "he stands still and runs to the door" — mutually exclusive instructions produce jittery, broken results. Reread the prompt and make sure camera motion and subject motion don't fight each other.
- No motion at all. The prompt only describes appearance, so the result is nearly a freeze-frame. Add at least one moving element: a subject action, a camera move, or a live environment (wind, smoke, water).
- Style drifting across a series. If clips meant for one edit don't match, the "scene passport" (character, light, style) diverged between prompts. Keep those blocks word-for-word identical.
Open Kling v3 in GPTunneL, start with one of the prompts above at 720p, nail the motion — then render the final version at 1080p.