Trim the Reference Clip Before You Upload It

Take a simple example: a fourteen-second phone clip of someone pouring coffee.

The useful movement lasts less than three seconds. The rest contains camera shake, background music, a face reflected in the window, and someone reaching for a cup.

Before uploading the recording to a reference-based tool such as the MiniMax H3 AI Video Generator, cut out the exact movement you need. More footage does not always provide clearer direction.

Why Trim the Clip First?

One recording can contain several actions, camera movements, voices, and lighting changes.

If the hand tilting the kettle is the only useful part, everything else may distract from that movement. A shorter clip also takes less time to review and makes unwanted faces, voices, logos, or private details easier to spot.

The aim is not to create a polished edit. It is to prepare a clear reference.

Find and Cut the Useful Motion

Check the source file before editing:

ffprobe -v error \
  -show_entries format=duration:stream=width,height,r_frame_rate \
  -of default=noprint_wrappers=1 \
  phone-coffee.mp4

Watch the recording and note where the action starts and ends:

start: 00:00:04.200
end:   00:00:06.800

Keep a few calm frames before the kettle moves and after it returns to the table. They show the starting and finishing positions.

The selected action lasts 2.6 seconds:

ffmpeg -i phone-coffee.mp4 \
  -ss 00:00:04.200 \
  -t 2.600 \
  -c:v libx264 \
  -crf 18 \
  -preset medium \
  -pix_fmt yuv420p \
  -movflags +faststart \
  -an \
  pour_04.2-06.8_noaudio.mp4

The main options are:

• -ss: starting point
• -t: output duration
• -an: remove the original audio
• -crf 18: create a high-quality re-encode
• -pix_fmt yuv420p: improve playback and upload compatibility

A lower CRF generally means higher quality and a larger file. For a short reference clip, 18 is a practical starting point rather than a required value.

Because -ss appears after the input, FFmpeg decodes toward the selected point. This is slower than a quick stream copy but usually produces a more accurate cut.

If the original sound matters and is cleared for use, remove -an and add:

-c:a aac

Otherwise, leave the audio out. Background music or conversation may add rights and privacy concerns without helping the movement.

Compare Two Cuts and Define Their Role

Make one version containing only the action and another with a little more preparation:

pour_04.2-06.8_noaudio.mp4
pour_03.7-06.8_noaudio.mp4

Check both clips:

• Is the starting pose visible?
• Does the movement finish naturally?
• Does the camera remain steady?
• Is another person entering the frame?
• Are reflections, screens, logos, or license plates visible?
• Is any original audio necessary?

The cleaner clip is usually more useful than the more dramatic one.

Add a short instruction explaining what the reference is meant to provide:

Use the reference only for the hand and kettle movement. Keep the camera still and maintain the cup position. Do not copy the background or reflected person.

This is guidance, not a guarantee that every other detail will remain unchanged. Keep the first test narrow so that problems are easier to identify.

Review the Whole Result

Do not check only whether the pouring motion looks similar. Watch the complete output for:

• A kettle that changes shape
• A hand that switches position
• A cup that moves before contact
• An action that starts or ends too early
• New objects appearing in the background
• Sound that does not match the visible action

Timing problems can often be fixed in a conventional editor. If the main object or movement changes across several moments, another generation may be more practical.

Check the Source Before Uploading

Use only footage you have permission to process. Remove unnecessary faces, voices, private screens, client material, and unreleased products.

Generated product demonstrations should not be treated as evidence of performance or safety. Verify important claims and physical instructions outside the generated clip.

Supported inputs and usage terms can differ between services and may change. Check the current interface and terms before using sensitive or commercial material.

The best reference is not the longest clip.

It is the clip that shows the intended movement—and little else.