Skip to main content

Tips & Best Practices

Things that save time and produce better results.

Run Analysis First​

Analysis is the slow part. Run it once and cache the results:

analysis:
  use_scene_detection: true

Subsequent runs load from cache instantly. This matters a lot when you're iterating on clip selection.

Use Hardware Acceleration​

If you have a GPU, use it. The tool auto-detects NVIDIA (NVENC), Apple (VideoToolbox), Intel (QSV), and AMD (VAAPI). Check what's available:

immich-memories hardware

Encoding a 10-minute 1080p video takes ~30 seconds with NVENC vs ~5 minutes on CPU.

Adjust Scene Detection Threshold​

The default threshold (27.0) works for most content, but you might need to tune it:

  • Lower threshold (e.g., 20.0) = more scene cuts detected. Good for fast-paced content with lots of action.
  • Higher threshold (e.g., 35.0) = fewer cuts. Better for slow, steady footage like landscapes.
analysis:
  scene_threshold: 27.0

Start with Shorter Durations​

Your first video should be 3-5 minutes, not 30. Shorter durations mean:

  • Faster generation
  • Easier to review
  • Less wasted time if your settings are off

Once you're happy with the results, scale up.

Review Clips Before Generating​

Step 2 exists for a reason. Spend 2 minutes deselecting clips that don't belong: that shaky hallway video, the accidental recording of your pocket, the 45-second clip of a wall. The tool's scoring is good but not perfect.

Enable LLM Analysis for Large Libraries​

For libraries with hundreds of videos, LLM content analysis makes a real difference in clip selection. It adds a few seconds per video to analysis time but catches things that motion/face detection misses: a quiet but meaningful conversation, a funny reaction shot, etc.

content_analysis:
  enabled: true
  weight: 0.35

Downscaling for Analysis​

If analysis is slow, enable downscaling. The tool doesn't need full-resolution frames to detect scenes and score clips:

analysis:
  enable_downscaling: true
  analysis_resolution: 480

This can cut analysis time in half with virtually no impact on clip selection quality.