AI Video Prompt Guide: Structure, Camera Controls & Model Differences

The 5-Part Prompt Formula That Works Across Models

Write every prompt in this order: Subject + Action + Camera + Lighting/Atmosphere + Style/Technical. This structure mirrors how diffusion models and video transformers parse tokens, reducing hallucinated limbs and temporal flicker. For example: "Cinematic close-up, elderly watchmaker assembling micro-gear, slow push-in on macro lens, warm tungsten key light, shallow depth of field, 8k, 24fps."

Subject & Action: Be Specific About Motion

Vague verbs like "moving" or "walking" produce sliding feet. Use precise motion descriptors:

• Camera-relative: push-in, pull-out, truck left/right, pedestal up/down, orbit, dolly zoom.
• Subject-relative: "turns head 15 degrees left," "raises right hand to chin," "fabric ripples in crosswind."
• Physics cues: "hair settles after spin," "water splashes upward then falls," "shadow shortens as sun rises."

Runway Gen-3 Alpha responds best when motion is described in the first 12 tokens; Kling 1.6 prefers motion tags at the end. Test both positions.

Camera Language: Translate Cinematography Into Tokens

Models trained on stock footage understand standard camera syntax. Include these four elements every time:

1. Shot type: extreme close-up, close-up, medium, wide, extreme wide, POV, over-the-shoulder.
2. Lens/focal length: 14mm (wide distortion), 24mm (environmental), 35mm (natural), 50mm (standard), 85mm (portrait compression), 200mm (telephoto flattening).
3. Aperture/depth: f/1.8 (shallow), f/2.8 (subject separation), f/5.6 (context), f/11 (deep focus).
4. Movement: static, handheld shake, smooth gimbal, crane up, slider lateral, FPV dive.

Example: "Medium shot, 35mm lens, f/2.8, slow gimbal truck right following subject."

Lighting & Atmosphere: Control Mood Without Post

Lighting tokens steer latent space toward consistent exposure. Specify:

• Source: golden hour sun, overcast sky, neon sign, practical lamp, bi-color LED panel, candle.
• Quality: hard (sharp shadows), soft (wraparound), diffused (silk), dappled (through leaves).
• Direction: key 45° camera-right, rim backlight, under-lighting (horror), top-down (interrogation).
• Color temp: 2700K warm, 5600K daylight, 6500K cool, mixed (neon cyan/orange).

Atmosphere tokens like "volumetric fog," "dust motes in shaft," or "heat haze" add depth but increase compute time by ~15% on Sora.

Model Comparison: Where Each Excels

Negative Prompting & Parameter Tuning

Most UIs expose CFG scale (guidance) and motion bucket (motion intensity). Start here:

• CFG 7–9: Balanced adherence vs. creativity. Above 10 causes burned highlights; below 5 ignores prompt.
• Motion bucket 40–60 (Gen-3) / 0.3–0.5 (Kling): Natural movement. Higher values create warp; lower freezes subject.
• Negative prompt defaults: "blur, low quality, distorted anatomy, extra limbs, watermark, text, title, credits, static, frozen, jitter, flicker, morphing, bad composition, oversaturated."

For Sora, use the style preset dropdown (cinematic, documentary, anime, etc.) instead of stuffing style tokens into the prompt.

Image-to-Video: Anchoring Consistency

When using I2V (Runway, Luma, Kling, Pika), the image provides 70% of the signal. Your prompt only needs to describe what changes:

1. Upload a reference frame (16:9, 1280x720 minimum).
2. Prompt only the delta: "Camera orbits left 30 degrees, subject smiles, wind lifts hair."
3. Set motion bucket lower (20–30) to prevent the model from "reimagining" the static elements.
4. Generate 4–6 variations, pick the cleanest, then extend 4 seconds at a time.

Common Failure Modes & Fixes

• Morphing hands: Add "hands resting on table," "holding coffee cup," or "wearing gloves" to constrain geometry.
• Sliding feet: Specify "weight shift left foot to right foot," "heel strike then toe-off."
• Flickering textures: Add "temporal consistency," "stable texture," or reduce CFG by 1.
• Wrong aspect ratio: Prefix with "16:9 cinematic," "9:16 vertical," "1:1 square."
• Ignored camera move: Move camera tokens to the very start (Gen-3) or very end (Kling) of prompt.

Workflow: From Idea to Usable Clip in 5 Minutes

1. Write the 5-part prompt in a text editor.
2. Generate 3 seeds at low resolution (512x288) to check composition.
3. Pick best seed, upscale to 1080p with motion bucket +10.
4. If >10 sec needed, use last frame as I2V input for next segment.
5. Batch color-grade in DaVinci Resolve: apply same LUT to all clips for visual cohesion.

FAQ

How many words should a prompt be?

40–80 tokens (roughly 30–60 English words). Longer prompts dilute attention; shorter ones under-specify.

Do I need to mention "4k" or "8k"?

Yes for Sora and Kling; they have resolution heads. Runway Gen-3 ignores resolution tokens—output is fixed at 1280x768.

Can I use Chinese prompts with Kling?

Kling 1.6 understands Chinese natively and often produces more accurate motion for Asian subjects with Chinese prompts. English works but may miss cultural nuance.

Why does my character's shirt change pattern mid-clip?

Texture sticking requires temporal attention. Add "consistent clothing texture," lower motion bucket, or use I2V with a locked reference frame.