SAM 3 Rotoscoping

Professional-grade background removal and object isolation using Meta's SAM 3

Tutorial

How It Works

1. Import video - Use Import Clip button or drag & drop

2. Click "Edit" - On video thumbnail

3. Select SAM 3 - From model selector (default for videos)

4. Choose selection method - Text prompt, box, or point selection

5. Track Frame - Preview mask on first frame (optional)

6. Track Entire Video - Process all frames

7. Remove Background - Export transparent ProRes 4444 video

Selection Methods

Text Prompt Selection

Best for: Clear subjects, simple scenes

1. Select "Text" input mode

2. Type description: "the person in the white shirt"

3. Click "Track Frame" - Preview on first frame

4. Review mask - Ensure correct selection

5. Track Entire Video - Process all frames

Example prompts:

• "the person"

• "the car"

• "the product"

• "the dancer in the center"

Box Selection

Best for: Precise control, complex scenes

1. Select "Box" input mode

2. Draw rectangle around subject

3. Optional: Add helper text describing selection

4. Click "Track Frame" - Preview mask

5. Track Entire Video - Process all frames

Tips:

• Draw box tightly around subject

• Include entire subject in box

• Helper text improves accuracy

Point Selection

Best for: Quick selection, single objects

1. Select "Point" input mode

2. Click on subject to add points

3. Add multiple points for better accuracy

4. Optional: Add helper text

5. Click "Track Frame" - Preview

6. Track Entire Video - Process

Tips:

• Click on center of subject

• Multiple points improve tracking

• Use for objects, not complex scenes

Workflow Steps

Step 1: Track Frame (Preview)

Purpose: Preview your selection before processing entire video

1. Make selection (text, box, or point)

2. Click "Track Frame"

3. Wait for processing (few seconds)

4. Review mask overlay on first frame

5. Adjust if needed - Change selection and track again

Why this matters:

• Saves time and credits

• Ensures correct selection

• Allows refinement before full processing

• Shows mask quality

Step 2: Track Entire Video

Purpose: Process all frames of your video

1. Confirm selection looks good from Track Frame

2. Click "Track Entire Video"

3. Wait for processing (scales with video length)

4. Progress indicator shows status

5. Mask data generated for all frames

Processing time:

• 10-second clip: ~30-60 seconds

• 30-second clip: ~2-3 minutes

• Longer clips take proportionally longer

Step 3: Remove Background

Purpose: Export transparent video file

1. After tracking completes

2. Click "Remove Background"

3. FFmpeg conversion creates transparent video

4. ProRes 4444 output with alpha channel

5. Download or Import to Premiere Pro

Output format:

• Format: ProRes 4444 (.mov)

• Alpha channel: Full transparency

• Quality: Professional grade

• Compatibility: Premiere Pro ready

Best Practices

Video Preparation

1. Keep clips short - Under 30 seconds for faster processing

2. Good contrast - Subject should stand out from background

3. Stable footage - Works best with steady shots

4. Clear subject - Well-defined subjects track better

5. Good lighting - Well-lit subjects produce cleaner masks

Selection Tips

1. Be specific in text prompts - "the person" vs. "person in blue shirt"

2. Draw tight boxes - Include entire subject, minimal background

3. Use multiple points - For point selection, add 2-3 points

4. Preview first - Always use Track Frame before full processing

5. Refine if needed - Adjust selection and re-track if mask is wrong

Processing Tips

1. Start with short clips - Test workflow with 5-10 second clips

2. Check first frame - Ensure Track Frame looks good

3. Be patient - Full video processing takes time

4. Monitor progress - Watch progress indicator

5. Save results - Download or import when complete

Use Cases

Product Isolation

Example:

• Import product demo video

• Text prompt: "the product"

• Track Entire Video

• Remove Background

• Result: Product on transparent background for compositing

Portrait Background Removal

Example:

• Import talking head video

• Box selection around person

• Track Entire Video

• Remove Background

• Result: Person isolated for new background

Object Extraction

Example:

• Import video with multiple objects

• Text prompt: "the car in the foreground"

• Track Entire Video

• Remove Background

• Result: Car isolated for compositing

Output Details

ProRes 4444 Format

Characteristics:

• Codec: Apple ProRes 4444

• Alpha channel: Full transparency support

• Quality: Professional, lossless alpha

• File size: Larger than MP4 (high quality)

• Compatibility: Premiere Pro, After Effects, Final Cut Pro

Importing to Premiere Pro

1. Download transparent video

2. Import into Premiere Pro project

3. Place on timeline above background

4. Composite with other footage

5. Alpha channel automatically used

Limitations

Video Length

• Recommended: Under 30 seconds

• Maximum: 5 minutes

• Processing time: Scales with length

• Longer clips: Consider trimming first

Resolution

• Maximum: 2160p (4K) for Kling O1 Edit

• Standard: 1080p recommended

• Higher resolution: Slower processing

• Lower resolution: Faster processing

Subject Complexity

• Simple subjects: Work best (single person, object)

• Complex scenes: May require refinement

• Multiple subjects: May need separate passes

• Overlapping objects: Can be challenging

Troubleshooting

"Track Frame not working"

Solutions:

• Ensure you've made a selection (text, box, or point)

• Check video is properly loaded

• Try different selection method

• Verify API keys are configured

"Mask is wrong on some frames"

Solutions:

• Use more specific text prompt

• Try box selection for better control

• Add helper text for box/point selection

• Process in shorter segments if needed

"Processing is too slow"

Solutions:

• Trim video to shorter duration

• Use lower resolution source

• Keep clips under 30 seconds

• Check internet connection

"Background not fully removed"

Solutions:

• Refine selection with Track Frame

• Use box selection for precise control

• Try different selection method

• Process again with adjusted selection

---

Next: Learn about Reshoot for video regeneration and style transfer.