Video Generation Best Practices

Master the art of AI video generation with proven techniques for selecting models, writing effective prompts, and optimizing your workflow for professional results.

---

Prepare Your Input Images

Start with properly sized source images for optimal generation quality.

Resolution Matters

Recommended workflow:

Why this matters:

• Models perform internal downscaling that reduces quality
• You maintain better control over composition and detail
• Generation quality improves when using optimally sized inputs

---

Choose the Right Model

Different models excel at different tasks. Select based on your specific needs.

Seedance 2.5 — create, edit, and extend

For workspaces with access, Seedance 2.5 is the strongest SeeDance option when you need:

• Longer output — up to 30 seconds (vs 15s on Seedance 2.0)
• Heavy reference stacks — up to 30 images, 10 videos, and 10 audio clips
• Three tasks — Create new video, Edit video on an existing clip, or Extend video forward/backward

Edit and Extend require a source video as Video 1, Adaptive aspect ratio, and (for Edit) Automatic duration. See the dedicated guide for setup and prompt examples.

Best Models Without Audio

Best Models With Audio

Model Selection Quick Guide

No audio needed:

• Kling 01 (first choice - best quality)
• SeeDance Pro (alternative - beautiful imagery)

Audio required:

• Kling 2.6 (best motion and acting)
• Veo 3.1 at 720p (best lip-sync)
• SeeDance Pro 1.5 (end frame needed)

---

Understanding Clip Timing

AI video models have built-in duration constraints that affect your workflow.

Working with Minimum Durations

Practical implications:

• You may only use part of each generation in your final edit
• Start frame plus prompt often sufficient for short clips
• End frames help guide motion but don't guarantee exact duration

Reverse Generation Technique

When to use reverse generation:

• Button or switch activations
• Object pickups or placements
• Precise mechanical actions
• Transitions to specific end states

Workflow:

1. Set your desired end state as the start frame
2. Prompt for the reverse action
3. Generate the clip
4. Reverse the footage in your video editor

---

Write Effective Prompts

Structure your prompts to give models clear direction while maintaining creative quality.

Core Prompt Formula

Describe the action, camera, and aesthetic in your prompts.

Action Prompting

Keep it simple, then iterate:

Camera Movement

Reliable camera prompts:

• Static camera
• Camera slowly moves forward
• Slowly zoom out
• Dolly camera up
• Rotate camera around the character
• Camera pans left across the scene

Pro tip: Use an end frame to show the desired camera position instead of relying solely on directional prompts.

Aesthetic and Quality

Add cinematic quality to your prompts by including high detail, realistic, filmed on Arri Alexa Mini, 4k references.

Why include camera references:

• Improves overall generation quality
• Guides the aesthetic toward cinematic results
• Helps maintain consistency across shots

Audio in Prompts

Audio prompt example:

The wizard smiles and speaks with a deep voice into the remote and says Sorcerer defying reality

Key elements:

• Describe voice characteristics like deep, soft, excited
• Use quotation marks around exact dialogue
• Include context such as speaking into, whispers to, shouts at

---

Advanced JSON Prompting

Take control of timing and action with structured JSON prompts for precise direction.

What is JSON Prompting

Benefits:

• Control action timing throughout the clip
• Specify multiple sequential actions
• More predictable results for complex scenes
• Professional-level control

How to Create JSON Prompts

Simple workflow:

1. Write your prompt as you normally would
2. Ask ChatGPT to convert your prompt into JSON format for video generation
3. Specify the clip duration when making the request
4. Review the generated JSON structure
5. Copy and paste the JSON into your prompt field

JSON Example

Example JSON structure for a 5-second clip with character standing, smiling and raising hand, then waving. Include timestamp markers for each action change and camera movement descriptions. Finish with aesthetic references like filmed on Arri Alexa Mini, cinematic, 4k.

When to use JSON:

• Complex multi-action sequences
• When precise timing matters
• Professional projects requiring repeatability
• Coordinating camera moves with actions

---

Unlimited Mode Strategy

Save credits during iteration by using unlimited mode when available.

Understanding Unlimited Mode

How to use it:

1. Check whether your plan and chosen model support draft or unlimited mode (shown in the generator when available)
2. Verify the mode supports the resolution you need for testing
3. Use draft/unlimited mode for iteration and prompt testing
4. Switch to standard credit-based generation for final high-quality outputs

Perfect for:

• Testing multiple prompt variations
• Iterating on camera movements
• Experimenting with different approaches
• Learning what works before committing credits

---

Iteration and Refinement

Achieve perfect results through strategic iteration.

Expect Multiple Generations

Typical iteration count: 5 to 20 generations per clip

Variables to refine:

• Prompt wording and detail level
• Input image composition
• AI model selection
• Camera movement description
• Start and end frame combination

Systematic Refinement

---

Lip-Sync and Voice Work

Handle dialogue and voice work with the right tools and techniques.

Best Models for Dialogue

Priority order:

1. Kling 2.6 for best overall audio quality and acting
2. Veo 3.1 for best lip-sync accuracy
3. SeeDance Pro 1.5 as decent alternative

When Lip-Sync Fails

Options for fixing poor lip-sync:

---

Quick Reference Guide

Model Selection Cheat Sheet

No audio:

• Kling 01 for best overall quality with end frame support
• SeeDance Pro for beautiful imagery with start frame only

With audio:

• Kling 2.6 for best acting and motion, no end frame, 1080p
• Veo 3.1 for best lip-sync, 720p only (upscale later)
• SeeDance Pro 1.5 for end frame support, 720p

Prompt Template

Action description, camera movement, high detail, realistic, filmed on Arri Alexa Mini, 4k

Example: The astronaut turns and waves at the camera, camera slowly dollies forward, high detail, realistic, filmed on Arri Alexa Mini, 4k

Pre-Generation Checklist

Before generating, verify:

• Images are 1920x1080 HD
• Model selected based on audio needs
• Prompt describes action plus camera plus aesthetic
• Audio dialogue in quotation marks
• Unlimited mode checked if iterating
• Ready to generate 5 to 20 variations

---

Troubleshooting

Poor Generation Quality

Problem: Blurry or low-quality video output

Solutions:

• Verify input images are 1920x1080
• Try Kling 01 for best quality
• Add high detail, 4k, filmed on Arri Alexa Mini to prompt
• Ensure start frame is well-composed and sharp

Wrong Motion or Action

Problem: Model doesn't follow your action prompt

Solutions:

• Simplify prompt to basic action only
• Try reverse generation technique
• Use end frame to show desired result
• Test with JSON prompting for complex actions

Camera Movement Issues

Problem: Camera doesn't move as prompted

Solutions:

• Use end frame instead of directional prompts
• Simplify camera description (just say camera moves forward)
• Try static camera if movement isn't critical
• Generate multiple variations since camera moves are inconsistent

Lip-Sync Problems

Problem: Audio and mouth movements don't match

Solutions:

• Try Veo 3.1 for best lip-sync
• Use ElevenLabs to adjust voice to match video
• Try sync.so lip-sync replacement tool
• Regenerate with simpler dialogue prompt

---

Next Steps

Recommended Learning Path

Related Resources

Seedance 2.5: Create, Edit & Extend

Learn Effective Prompt Writing

Common Issues & Solutions

---