Replace a person in a video and keep the original audio

What actually changes in a character replacement?

A whole-character replacement aims to reinterpret the visible performer using a reference image. That can affect the face, clothing, silhouette, and other visual details. It is more ambitious than changing a face region, so do not assume every background pixel or gesture will survive unchanged.

Audio does not need to change just because the picture changes. Your source may contain speech, music, room tone, or a combination of them. Preserving that track means carrying its existing sound into the new video; it does not mean reconstructing the voice or making the new character speak differently.

Start by deciding what must remain consistent. For a dance clip, musical timing may be the priority. For a mascot waving, the shape of the gesture matters more than exact facial movement. For dialogue, mouth timing becomes an additional quality check. This decision helps you reject an unsuitable source before spending credits.

Prepare one clean, short source clip

Use a single shot with one clearly visible adult performer. A steady camera, moderate movement, and separation between the subject and background give you fewer details to inspect later. If the subject repeatedly passes behind furniture or leaves the frame, start with an easier passage.

The TryMotion web studio accepts a prepared H.264 MP4 between 1 and 10 seconds, at up to 30 frames per second, with a maximum upload size of 25 MB. Trim it in your usual editor before uploading. There is no built-in trimming tool, and renaming a MOV file to MP4 will not convert its codec.

Listen to the trimmed file all the way through. Its audio should start at the intended moment, without a leftover lead-in from a longer recording. If your editor has moved the sound independently of the picture, fix the source timeline now. An audio preservation step cannot recover a timing relationship that was already wrong in the export.

Choose a reference that fits the shot

Your reference supplies the character's visible identity, not an unlimited description of hidden details. A headshot gives little guidance for a full-body performance. When the source shows the whole performer, use a reference that shows the whole character with a readable outline.

Match the pose and viewing direction as closely as practical. A front-facing reference paired with a source that stays in profile creates a harder transformation. Fine lettering, small logos, fingers, and accessories are details to inspect, not details to assume the model will reproduce faithfully.

Keep a copy of the reference used for each take. If you later compare two outputs, changing only the source or only the reference gives you a more useful comparison than changing everything at once. A simple record of the shot, reference, and intended gesture is enough to begin.

Keep audio separate from visual generation

A reliable workflow produces the replacement picture first and then combines that picture with the source soundtrack. The picture and sound should share the same starting point. The final export must contain an audio track, rather than merely showing an enabled sound icon in a preview.

TryMotion's implementation includes an optional audio merge stage when the uploaded video contains audio. The studio detects whether an audio track exists and includes the setting in the generation quote. If the source is silent, turning on preservation cannot create missing sound.

This workflow uses a short source clip and an original character reference. P-Video-Replace 720p is the default, with 1080p available for higher quality. Illustrated scenes describe the workflow and are not evidence of a particular generated result.

Review these checks before sharing

Watch the output at normal speed first, then inspect the difficult moments. A visually attractive first frame can conceal a drifting hand, a changing costume, or a lost face halfway through the shot. Review the entire clip rather than relying on its thumbnail.

1. Check the opening and closing frames. Confirm that the clip has the expected start, end, framing, and orientation.
2. Listen through with headphones. Verify that the track is present, audible, and free of an unexpected second soundtrack.
3. Look for a visible timing cue, such as a clap, footfall, or spoken consonant. Compare it with the source.
4. Inspect hands, face boundaries, occluded body parts, and any brand marks. Make sure the output is suitable for your purpose.
5. Download and replay the exported file in a second player. Confirm that the saved video includes the audio, rather than depending on a browser preview.

Troubleshoot the most common mismatches

If the soundtrack is absent, first check whether the uploaded source actually has an audio stream. A silent export can look identical to a normal MP4 in a file browser. If the sound exists locally but not in the final file, the issue belongs to the merge or export stage rather than the character reference.

If the soundtrack is consistently early or late, inspect the source edit and the output's first frame. A fixed offset can often be corrected in a video editor, but do not hide a recurring pipeline offset: it should be investigated before generating a batch. If synchronization drifts gradually, check whether the picture has been retimed or its duration changed.

If the voice sounds different, a generated audio track may have been used instead of the original recording. Preserve the source track when identity and performance matter. Voice cloning, dubbing, and speech synthesis should be separate, explicit choices with the right permissions; they are outside this MVP.

If the character breaks during a fast turn, simplify the picture problem. Try a shorter section, a steadier shot, or a reference closer to the source angle. Repeatedly running the same difficult input is not a reliable strategy for controlling cost.

Understand credits and retention

Default P-Video-Replace 720p costs 10 credits for a source up to 5 seconds and 20 for up to 10 seconds. The higher-quality 1080p option costs 20 and 40 credits respectively. The server checks the uploaded video's duration before issuing a quote; review that confirmed price before submitting.

Credits are reserved when a generation starts. A confirmed technical failure automatically returns the full reserved credits once. Stuck or unconfirmed jobs have a 30-minute overall limit. Saving a completed output is retried for up to 10 minutes, within that overall limit. If the video still cannot be delivered, credits return automatically on the next recovery check, normally within 5 minutes, even if you close the page. No support request is required. Unconfirmed submissions are checked without automatically submitting a second paid generation. A completed video that you dislike creatively is different from a job that technically failed.

TryMotion stores uploads and outputs privately with a 7-day expiry. Download the result you want to keep. Deleting a TryMotion file removes that stored copy, but inference providers may retain submitted inputs and outputs under their own policies. Avoid uploading material that you are not authorized to share for processing.

A practical first experiment

Begin with a 5-second wave, turn, or simple step performed against an uncluttered background. Use an original character reference, keep its entire silhouette visible, and preserve a short ambient soundtrack. Write down three acceptance criteria before you generate: the gesture remains recognizable, the character stays coherent, and the audio starts at the correct moment.

Review the result against those criteria before trying a more complicated shot. This gives you a concrete way to decide whether the workflow fits your project. The goal of a first test is to learn what the source and reference can support, then use that information to plan the next take.