Reference Mode

Generate videos with consistent characters using 1-9 reference images. Perfect for maintaining character appearance across multiple shots, creating talking head videos, or ensuring visual consistency

Attaching exactly 2 images and seeing Transition Mode? By default, 2 uploaded images trigger Transition Mode (start frame → end frame morph). To use them as character reference photos instead, select a dedicated reference model from the model picker: Kling O3 Reference, Seedance 2 Reference, or Wan 2.7 Reference. These models treat multiple images as reference material, not as start/end frames. See Transition Mode if you actually want a morph.

How It Works

1. Enable Generate Media - Toggle the Generate Media button in the composer

2. Attach reference images - Upload 1-9 images of the same character/subject (count depends on model)

3. Reference Mode activates - System detects multiple images and switches to reference model

4. Select reference model - Veo 3.1 Reference, Kling O3, Wan 2.7 Reference, or Seedance 2 Reference (auto-selected)

5. Describe the video - Write a prompt describing the scene and action

6. Configure settings - Set duration (varies by model: 8s for Veo, 3-15s for Kling O3, 2-10s for Wan, 4-15s for Seedance)

7. Generate - Click send to create a video with consistent character

Automatic Activation

Reference Mode activates when:

• 1-9 images attached - Reference images detected (minimum count varies by model)

• No videos attached - Reference mode is for generation, not editing

• Model auto-switches - Compatible reference models appear

• Element tags - Using @Element tags also triggers reference mode

Supported Models

Model	Reference Images	Duration	Audio	Best For
Veo 3.1 Reference	2-4 images	8s (fixed)	✅	Character consistency with audio
Kling O3	2-7 images	3-15s	✅	Maximum reference images, audio support, best quality
Wan 2.7 Reference	1-9 images	2-10s	❌	Most reference images, high resolution
Seedance 2 Reference	Up to 9 images	4-15s	✅	Reference with native audio

Reference Image Guidelines

Image Requirements

What makes good reference images:

1. Same subject - All images should show the same person/character

2. Clear face/features - Good visibility of key characteristics

3. Varied angles - Different views help the model understand the subject

4. Consistent lighting - Similar lighting conditions work best

5. High quality - Clear, well-lit images produce better results

Image Count Strategy

1-3 images:

• Minimum for reference mode (1 image minimum with Wan 2.7 Reference)

• Works well for simple characters

• Faster processing

4-7 images:

• Maximum consistency with Kling O3

• Better for complex characters

• More reference points for the model

Up to 9 images (Wan 2.7 Reference / Seedance 2 Reference):

• Highest number of reference inputs

• Best for complex subjects requiring many angles

• Wan 2.7 Reference supports 1-9 images; Seedance 2 Reference supports up to 9

What to Include

✅ Good Reference Images:

• Front-facing portrait

• Side profile

• 3/4 angle

• Different expressions

• Various lighting conditions

• Different outfits (same person)

❌ Poor Reference Images:

• Different people

• Unclear/blurry faces

• Extreme angles

• Very different styles

• Inconsistent subject

Writing Reference Mode Prompts

What to Describe

Focus on the scene and action, not the character (reference images handle that):

1. Setting - Where the scene takes place

2. Action - What the character is doing

3. Camera movement - How it's shot

4. Style - Visual aesthetic and mood

5. Dialogue (Veo only) - What the character says

Good Reference Mode Prompts

✅ With Action:

The character walks confidently through a modern office, greeting 
colleagues with a warm smile, a professional setting, and natural lighting

✅ With Dialogue (Veo):

The character speaks directly to the camera: "Welcome to our presentation. 
Today we'll explore innovative solutions." Professional setting, 
confident delivery

✅ With Camera Movement:

Slow push-in on the character as they explain a concept, cinematic 
style, soft lighting, engaging and professional atmosphere

Bad Reference Mode Prompts

❌ Describes Character:

A person with brown hair and blue eyes

(Reference images already provide this)

❌ Too Vague:

The character does something

(Not specific enough)

❌ Missing Context:

Person talking

(Needs setting, style, camera work)

Use Cases

Talking Head Videos

Example:

• 3-4 reference images of the speaker

• Prompt: "The speaker addresses the camera with confidence, explaining key concepts in a professional setting, and with natural lighting."

• Model: Veo 3.1 Reference (for audio)

• Result: Consistent talking head video with audio

Character Consistency

Example:

• 4-5 reference images of a character

• Prompt: "The character walks through a futuristic city, looking around with curiosity, cinematic style, neon lighting."

• Model: Kling O3 (for quality)

• Result: Character maintains appearance across shots

Product Demonstrations

Example:

• 2-3 reference images of a presenter

• Prompt: "The presenter demonstrates a product, showing features with enthusiasm, bright studio lighting, professional setting."

• Model: Veo 3.1 Reference (for audio)

• Result: Consistent presenter across multiple shots

Narrative Scenes

Example:

• 5-7 reference images of the main character

• Prompt: "The character enters a mysterious room, the camera follows their gaze, cinematic style, moody lighting, suspenseful atmosphere."

• Model: Kling O3 (for quality and longer duration)

• Result: Consistent character in narrative scene

Model-Specific Behavior

Veo 3.1 Reference

Characteristics:

• Reference images: 2-4

• Duration: 8 seconds (fixed)

• Aspect ratio: 16:9 or 9:16 only (auto-determined from reference images)

• Audio: Supported (dialogue generation)

• Resolution: 720p-1080p

Best for:

• Talking head videos

• When you need audio/dialogue

• Character consistency with speech

• Professional presentations

Limitations:

• Fixed 8-second duration

• Maximum 4 reference images

• Aspect ratio limited to 16:9 or 9:16 (auto-determined from images)

Kling O3

Characteristics:

• Reference images: 1-7

• Duration: 3-15 seconds (variable, Pro/Standard toggle)

• Aspect ratio: 16:9, 9:16, or 1:1

• Audio: Supported (native audio generation)

• Resolution: 720p

Best for:

• Maximum character consistency with audio

• Flexible duration (3-15 seconds)

• More reference images (up to 7)

• High-quality results with sound

Limitations:

• Fixed aspect ratios (no auto)

Wan 2.7 Reference

Characteristics:

• Reference images: 1-9

• Duration: 2-10 seconds (variable)

• Aspect ratio: 16:9, 9:16, 1:1, 4:3, or 3:4

• Audio: Not supported

• Resolution: 720p-1080p

Best for:

• Maximum number of reference images (up to 9)

• High resolution output

• When you need the most reference angles

• Flexible aspect ratios

Limitations:

• No audio generation

• Shorter maximum duration than other models

Seedance 2 Reference

Characteristics:

• Reference images: Up to 9

• Audio reference files: Up to 3 (attach audio to influence generated sound)

• Duration: 4-15 seconds (variable)

• Aspect ratio: Auto-detected from input images

• Audio: Supported (native audio generation)

• Resolution: 480p-1080p

Best for:

• Reference videos that need a native soundtrack

• Longer reference outputs (up to 15s)

• When you want audio generated automatically

• When providing reference audio to guide the sound design

Limitations:

• Audio may need refinement in post

Tips for Best Results

1. Use varied reference images - Different angles and expressions help

2. Keep images consistent - Same person, similar quality

3. Describe the scene, not the character - Reference images handle appearance

4. Use appropriate model - Veo for audio, Kling for more references

5. Match aspect ratios - Reference images should have similar ratios

6. Be specific about action - Describe what the character does

7. Include camera movement - Helps create dynamic videos

8. Test with different image counts - Find what works best for your character

Common Workflows

Quick Talking Head

1. Prepare 2-3 reference images

2. Attach images

3. Select Veo 3.1 Reference (auto-selected)

4. Prompt: "Character speaks to camera: '[dialogue]', professional setting"

5. Enable audio

6. Generate (8 seconds)

High-Quality Character Video

1. Prepare 5-7 reference images

2. Attach images

3. Select Kling O3 (auto-selected)

4. Detailed prompt with scene, action, and style

5. Generate (3-15 seconds)

Consistent Character Series

1. Prepare 4-5 reference images

2. Create multiple videos with different prompts

3. Character appearance stays consistent

4. Use for series or multiple shots

Reference with Audio (Seedance 2)

1. Prepare up to 9 reference images

2. Attach images

3. Select Seedance 2 Reference

4. Describe the scene, action, and desired audio

5. Audio generates automatically

6. Generate (4-15 seconds)

Troubleshooting

"Reference Mode not activating."

Solutions:

• Ensure you have 1-9 images attached

• Check that images are properly uploaded

• Remove any attached videos (reference mode is for generation)

• Verify you're in Generate Media mode

"Character doesn't look consistent."

Solutions:

• Use more reference images (4-7 for best results)

• Ensure all images show the same person

• Use varied angles and expressions

• Try Kling O3 for better consistency

• Check image quality (clear, well-lit)

"Model doesn't support reference mode."

Solutions:

• Switch to Veo 3.1 Reference, Kling O3, Wan 2.7 Reference, or Seedance 2 Reference

• Check Supported Video Models

• Most models don't support reference mode

• Reference mode requires specific models

"Duration is wrong."

Solutions:

• Veo 3.1 Reference: Fixed at 8 seconds (cannot change)

• Kling O3: 3-15 seconds (variable, set in settings)

• Wan 2.7 Reference: 2-10 seconds (variable, set in settings)

• Seedance 2 Reference: 4-15 seconds (variable, set in settings)

• Reference mode has limited duration control

• Plan your content for the available duration

---

Next: Learn about Image Generation for creating still images.