Prompt Guide

Master the art of writing effective prompts for AI video generation. Learn techniques, patterns, and best practices to get the best results.

Prompt Basics

A prompt is a text description of the video you want to generate. The AI interprets your words and creates a video that matches your description as closely as possible. Here are the fundamentals:

Key Principle: Be specific and descriptive. The more detail you provide about the scene, the better the AI can understand your vision. Vague prompts produce unpredictable results.
  • Be descriptive — Instead of "a dog", write "a golden retriever playing fetch in a sunlit park"
  • Specify the camera — Include camera angle and movement: "close-up", "wide shot", "tracking shot"
  • Define the mood — Use lighting and atmosphere terms: "golden hour", "moody", "dramatic"
  • Set the style — Specify the visual approach: "photorealistic", "cinematic", "anime", "watercolor"
  • Include motion — Describe what's moving: "waves crashing", "birds flying", "camera panning left"

Prompt Structure

Follow this proven structure for consistent, high-quality results:

1. Camera/Shot Type Cinematic tracking shot
2. Subject of a lone wolf walking through
3. Environment a snow-covered forest at dawn
4. Lighting & Atmosphere soft golden light filtering through mist
5. Style & Quality photorealistic, 4K, shallow depth of field
Complete Prompt
"Cinematic tracking shot of a lone wolf walking through a snow-covered forest at dawn, soft golden light filtering through mist, photorealistic, 4K, shallow depth of field"

Camera & Motion Keywords

Use these terms to control how your video looks and moves:

Shot Types

close-up extreme close-up medium shot wide shot establishing shot aerial shot bird's eye view low angle high angle dutch angle POV shot over-the-shoulder

Camera Movement

tracking shot dolly zoom pan left/right tilt up/down crane shot steadicam handheld orbiting push in pull back slow zoom whip pan

Motion Speed

slow motion time-lapse hyperlapse real-time speed ramp freeze frame

Style & Aesthetics

Visual Style

photorealistic cinematic film grain anamorphic documentary 3D render anime oil painting watercolor pixel art minimalist surrealist

Lighting

golden hour blue hour harsh sunlight soft light backlit rim lighting neon lighting volumetric studio lighting natural light candlelight moonlight

Mood & Atmosphere

dramatic ethereal moody dreamy gritty peaceful mysterious epic nostalgic futuristic

Negative Prompts

Negative prompts tell the AI what to avoid. They're essential for preventing common artifacts and quality issues.

Recommended Default Negative Prompt
blurry, distorted face, deformed eyes, extra limbs, extra fingers, mutated hands, bad anatomy, bad proportions, watermark, text, logo
Tip: The default negative prompt is pre-filled in the generator. You can customize it per-generation to avoid specific unwanted elements.

Common Negative Terms

blurry low quality distorted deformed watermark text overlay bad anatomy extra limbs jpeg artifacts oversaturated static flickering

Model-Specific Prompt Techniques

Different AI models respond to prompts differently. Understanding each model's strengths lets you write prompts that play to those strengths.

WAN 2.2 — Speed & Iteration

WAN is fast (4 steps) and excels at clear, simple compositions with natural motion. It responds well to direct, concrete language.

WAN — What Works
Close-up of a woman's face, wind blowing through her hair, golden hour light, shallow depth of field, photorealistic, cinematic
WAN Tips: Keep prompts focused on one subject with one action. WAN struggles with complex multi-subject scenes. Use it for rapid iteration — test 10 prompt variations in the time it takes another model to render once.

LTX 2.3 & LTX Quality — Detail & Resolution

LTX uses a two-pass pipeline that excels at sharp detail and HD resolution. It handles more complex scenes and responds well to technical language.

LTX — What Works
Tracking shot following a vintage motorcycle driving down a winding coastal road at sunset, ocean waves crashing against cliffs in the background, dust particles visible in the golden backlight, anamorphic lens, cinematic color grading, ultra-detailed
LTX Tips: Add detail qualifiers like "ultra-detailed", "4K", "sharp focus" — LTX's upsampler pass amplifies these. Use "anamorphic" or "shallow depth of field" for cinematic lens effects. LTX Quality adds a LoRA refinement pass, so technical prompts with many descriptors produce noticeably sharper output than on WAN.

Hunyuan & Hunyuan 1.5 — Motion & Characters

Hunyuan produces exceptionally smooth motion and handles character animation, anime, and complex prompt comprehension better than other models.

Hunyuan — What Works
A samurai warrior drawing his katana in slow motion, cherry blossom petals swirling in the wind around him, dramatic side lighting revealing the blade's reflection, anime-influenced style with photorealistic textures, epic cinematic
Hunyuan Tips: Hunyuan 1.5 has the best text comprehension, so long complex prompts work well. Describe character appearance in detail — "a middle-aged man with grey stubble, wearing a worn leather jacket" gives Hunyuan more to work with. Use CFG 6.0 (it's tuned for this). Mention specific anime studios or art styles for stylised output.

CogVideoX — Artistic & Stylised

CogVideoX excels at artistic, creative, and stylised output. It has particularly strong text comprehension from research-grade training.

CogVideoX — What Works
A surrealist painting coming to life, melting clocks dripping off the edge of a desert cliff, Salvador Dali style, dreamlike atmosphere, impossibly saturated colours, time flowing like liquid, artistic masterpiece

AnimateDiff — Anime & SD Styles

AnimateDiff turns Stable Diffusion checkpoints into animations. It's best for short, stylised clips — especially anime, illustration, and any SD-compatible art style.

AnimateDiff — What Works
1girl, long silver hair flowing in the wind, standing on a moonlit rooftop, cityscape background, petals floating, anime style, detailed eyes, soft glow, masterpiece, best quality
AnimateDiff Tips: Use Stable Diffusion prompt syntax: comma-separated tags work better than natural sentences. Include quality tags like "masterpiece, best quality". Keep motion subtle — "flowing hair", "drifting clouds", "floating particles" work well. Avoid fast camera movement.

Image-to-Video Prompt Guide

When generating Image-to-Video (I2V), your prompt describes what should happen to the static image, not what the image looks like. The model already has the image — it needs to know how to animate it.

I2V Prompt Structure

1. Subject Action The woman slowly turns her head to the right
2. Camera Movement camera slowly pushing in
3. Environmental Motion wind gently blowing through the trees in the background
4. Atmosphere Details golden dust particles floating in the light

I2V Do's and Don'ts

✓ Do

  • Describe motion: "the dog runs forward", "waves crash on rocks"
  • Specify camera: "camera slowly pans left", "gentle zoom in"
  • Add ambient motion: "clouds drifting", "leaves rustling"
  • Keep it simple: one or two actions per prompt
  • Match your image: if the image is a portrait, describe portrait motion

✗ Don't

  • Don't describe the scene: the model already sees the image
  • Don't request scene changes: "turns into a cat" won't work
  • Don't add conflicting motion: "zooming in while pulling back"
  • Don't use style keywords: style is locked by the source image
  • Don't write long complex prompts: I2V works best with short, focused motion descriptions

I2V Example Prompts

Portrait Photo
I2V Prompt
The person smiles gently and tilts their head slightly, hair moves softly with a breeze, subtle camera push in, soft focus on background
Landscape Photo
I2V Prompt
Clouds slowly drift across the sky, water ripples gently, grass sways in a light breeze, camera slowly pans right across the scene
Product Photo
I2V Prompt
Camera slowly orbits around the product, reflections shift on the surface, subtle depth of field animation, smooth studio turntable motion

Advanced Techniques

Temporal Keywords

Use time-based language to control transitions within a single clip:

Temporal Phrasing

transitioning from... to... gradually changing slowly revealing emerging from dissolving into morphing from... to... beginning with... ending with... unfolding over time
Temporal Example
Time-lapse of a flower bud gradually opening into full bloom, transitioning from dawn pink light to bright midday sun, petals slowly unfurling, dew drops evaporating, macro lens, nature documentary quality

Prompt Weighting & Emphasis

Place the most important elements at the beginning of your prompt. AI models weight earlier terms more heavily:

✗ Weak

"There is a forest with some fog and a deer walking through it in a cinematic way with golden light"

✓ Strong

"Cinematic tracking shot of a majestic deer walking through a foggy ancient forest, golden volumetric light rays, photorealistic, film grain"

Combining Motion Types

Layer different types of motion for cinematic depth:

  • Camera + Subject: "Tracking shot following a running athlete" — camera moves while subject moves
  • Camera + Environment: "Slow crane shot rising above a city as cars move below" — camera moves, environment has motion
  • Subject + Particles: "Dancer spinning as sparks and embers swirl around her" — subject motion with particle effects
  • All Three: "Dolly push through a busy marketplace, vendors calling out, banners fluttering in the breeze, dust motes in shafts of light" — camera, subjects, and environmental detail

Quality Boosters

Append these terms to almost any prompt to improve output quality:

General Quality

cinematic 4K 8K photorealistic ultra-detailed professional award-winning masterpiece

Lens & Film

shallow depth of field bokeh anamorphic film grain lens flare 35mm IMAX tilt-shift

Composition

rule of thirds leading lines symmetrical centred subject negative space golden ratio foreground interest framed composition

Common Mistakes

Avoid these pitfalls that lead to poor results:

Too Vague

"A nice video of nature" — gives the AI almost nothing to work with. Be specific: what nature? What time of day? What camera angle? What mood?

Too Many Subjects

"A man and woman dancing while a dog runs by and fireworks explode overhead and cars drive past" — too many competing elements. Focus on one clear subject and action per generation.

Contradictory Instructions

"Close-up aerial wide shot" or "bright dark moody sunny" — conflicting terms confuse the model and produce incoherent output.

High CFG with Low-CFG Models

Setting CFG to 7+ on WAN or LTX (which are tuned for CFG 1.0) produces oversaturated, distorted results. Trust the default CFG for each model.

Expecting Text Rendering

AI video models cannot reliably render readable text, logos, or specific written words in video. Avoid prompts like "a sign that reads 'Hello World'".

Ignoring Duration

A 1-second clip can't show complex action sequences. Match your prompt complexity to the video duration: simple motion for short clips, progressive scenes for longer ones.

Genre Templates

Ready-to-use prompt templates for common video genres. Copy, customise the bracketed sections, and generate.

🎥 Cinematic / Film
Template
Cinematic [shot type] of [subject] in [environment], [lighting description], [camera movement], film grain, anamorphic lens, shallow depth of field, professional color grading, 4K
👾 Horror / Thriller
Template
Eerie [shot type] of [subject] in [dark environment], dim flickering light casting long shadows, [slow unsettling camera movement], fog rolling across the ground, desaturated cold tones, tension building, horror film aesthetic
💙 Romance / Drama
Template
Intimate [shot type] of [subjects] in [warm environment], soft golden hour light, [gentle camera movement], bokeh background, warm colour palette, romantic atmosphere, shallow depth of field, cinematic
🏃 Action / Thriller
Template
Dynamic [tracking/handheld] shot of [subject in action] through [high-energy environment], [fast camera movement], motion blur on background, high contrast lighting, adrenaline-fueled atmosphere, cinematic, professional stunt photography
🌎 Documentary / Nature
Template
[Aerial/tracking] shot of [wildlife/landscape] in [natural habitat], [natural lighting], [steady camera movement], incredible detail, National Geographic quality, nature documentary, photorealistic, 4K ultra HD
🎵 Music Video
Template
[Creative shot type] of [performer/subject] in [stylised environment], [dramatic/colourful lighting with specific colours], [rhythmic camera movement], high fashion aesthetic, music video vibes, vivid colours, artistic composition
💼 Product / Commercial
Template
Smooth [orbiting/tracking] studio shot of [product] on [clean surface/background], [professional studio lighting], [slow elegant camera movement], reflections visible on surface, shallow depth of field, luxury commercial photography, high-end advertising
🔮 Sci-Fi / Fantasy
Template
Epic [shot type] of [subject] in [otherworldly environment], [volumetric/neon/ethereal lighting], [dramatic camera movement], particle effects, [futuristic/magical] atmosphere, ultra-detailed, concept art quality, 8K

Optimal Settings per Model

WAN 2.2 — Recommended Settings

Steps4 (default) — increase to 6-8 for more detail
CFG1.0 — keep low for natural results
SamplerEuler (two-pass KSamplerAdvanced)
SchedulerSimple
Shift5.0
Resolution848×480 for best results
Duration1-3 seconds for fast iteration

LTX 2.3 — Recommended Settings

Steps8 (default) — the two-pass pipeline handles quality
CFG1.0 — keep at default for both passes
SamplerEuler Ancestral CFG++ (required for this model)
SchedulerSimple
Shift2.05
Resolution1280×720 for full quality
Duration2-5 seconds for best results

LTX Quality — Recommended Settings

Steps20 (first pass), uses LoRA refinement
CFG3.0 (first pass) / 1.0 (upscale pass)
SamplerEuler
SchedulerLTXVScheduler (custom shift)
Shift0.95 – 2.05
ResolutionUp to 3840×2176 (4K)
Duration2-5 seconds for best detail

Hunyuan 1.5 — Recommended Settings

Steps20 — sweet spot for motion quality
CFG6.0 — higher than WAN/LTX, tuned for this
SamplerEuler
SchedulerSimple
Shift7.0
Resolution1280×720 (native HD)
Duration2-5 seconds for smooth motion

CogVideoX — Recommended Settings

Steps50 — needs more steps for quality
CFG6.0
SamplerEuler
Resolution720×480
Duration~6 seconds (49 frames at 8fps)

AnimateDiff — Recommended Settings

Steps20 — standard SD step count
CFG7.0 — standard SD guidance scale
SamplerEuler
Resolution512×512
Duration~4 seconds (32 frames at 8fps)

Example Prompts

Copy these prompts and try them out. Each is crafted to showcase different capabilities.

Nature & Landscape
Prompt
Cinematic aerial shot slowly descending over an ancient redwood forest shrouded in morning mist, shafts of golden sunlight breaking through the canopy, a gentle river winding through the valley below, photorealistic, National Geographic quality
Sci-Fi & Fantasy
Prompt
A colossal space station orbiting a gas giant planet, camera slowly pushing through docking bay windows to reveal thousands of tiny ships approaching, volumetric nebula lighting in purple and teal, hard sci-fi aesthetic, Kubrick-inspired composition
Product & Commercial
Prompt
Smooth 360-degree orbiting shot of a luxury wristwatch on a marble pedestal, dramatic studio lighting with soft shadows, reflective surfaces catching light, shallow depth of field with bokeh background, high-end commercial photography style
Abstract & Artistic
Prompt
Mesmerizing slow-motion macro shot of ink drops falling into clear water, splitting into fractal tendrils of deep blue and crimson, abstract fluid dynamics, studio lighting against pure black background, high speed photography
Urban & Architecture
Prompt
Hyperlapse tracking through Tokyo streets at night, neon signs reflecting on rain-slicked pavement, crowds of people moving at accelerated speed, camera weaving through narrow alleys, cyberpunk atmosphere, anamorphic lens flares

Pro Tips

1

Start short, refine long

Begin with 1-2 second clips using WAN to test your prompt concept. Once you're happy with the direction, switch to LTX and extend the duration for your final output.

2

Use seeds for consistency

When you find a result you like, note the seed value. Use the same seed with slight prompt variations to explore similar outputs without starting from scratch.

3

Specify camera first

Leading with the shot type and camera movement gives the model the strongest guidance for composition. "Tracking shot of..." works better than "...tracking shot".

4

Quality keywords matter

Adding "cinematic", "4K", "photorealistic", or "professional" at the end genuinely improves output quality. The model has learned associations between these words and higher quality.

5

Less is more with CFG

Keep CFG at 1.0 for both models. Higher values don't mean better results — they can cause over-saturation and artifacts. These models are calibrated for low CFG.

6

Match aspect to content

Use 16:9 for landscapes and cinematic shots, 9:16 for social media verticals, 1:1 for abstract art. The model understands composition differently per aspect ratio.

Ready to try these prompts?

Create a free account and paste any of these examples to see the results for yourself.