Mirelo SFX 2.0
Add synced sound effects, foley, and ambience to any video, or generate sound from a text prompt.
Mirelo SFX 2.0: Synced Sound Effects, Foley, and Ambience for Any Video
Upload a video and get it back with a soundtrack that matches what happens on screen. Mirelo SFX 2.0 watches the picture, places every footstep, door, engine, and gust where it belongs, and keeps the ambience steady across cuts. It keeps the dialogue, drops the music, and leaves the picture untouched.
No video? Describe the sound and it generates it from text, up to ten minutes long.
What can you do with it?
- Finish a silent AI clip. Generated video usually comes out mute or with a thin, generic track. Run it through once and the scene has weight: tires on gravel, rain on glass, a crowd that swells when the camera pans.
- Replace the audio on a clip you already have. Keep the spoken performance, strip the music you don’t have rights to, and get fresh effects and ambience woven around the voices.
- Score a sequence after the edit. Stitch your shots together first, then add sound to the whole cut. The ambience carries across cuts of the same location instead of resetting at every edit.
- Build a sound library from text. “Heavy rain on a metal roof with distant thunder.” “A subway car braking into a station.” Ask for several variations and pick the one that fits.
- Make off-screen sound to place yourself. Generate a sound effect from text and drop it on your timeline where you need it.
- Hear a scene four ways. Up to four variations per run, each a different take on the same footage.
How to use it
- Upload a video, or skip this and write a prompt to generate sound from text alone.
- Choose the output. Video with sound hands you the clip ready to post. Soundtrack only gives you the audio files to mix yourself.
- Decide what happens to dialogue. Keep it and the new sound is built around the voices. Drop it and you get effects and ambience only.
- Guide it, if you want. A short prompt steers the result: “wet asphalt, sirens far away, no music.” The model reads most of what it needs from the picture.
- Run it. Pick how many variations you want and generate.
What you get
For a video: your clip, same picture, new sound, as an mp4. Or, if you chose soundtrack only, one audio file per variation, cut to the length of the video so it drops onto your timeline in sync.
For a text prompt: one audio file per variation at the length you set, up to ten minutes.
Audio comes back as WAV by default. MP3, AAC, and FLAC are available in the settings.
Settings
- Output: the video with its new soundtrack, or the soundtrack on its own.
- Dialogue: keep the speech from the original audio, or remove it and generate effects only.
- Duration: for text prompts, how many seconds to generate. Videos are always scored end to end.
- Variations: one to four takes per run.
- Negative prompt: sounds to keep out, as a short comma-separated list.
- Seed: fix it to reproduce a take, or leave it empty for a fresh one.
- Audio format: WAV, MP3, AAC, or FLAC.
Tips
- Let the picture lead. For video, start with no prompt at all. Add a few words only when the result misses something you can’t see, like weather off-screen or a specific vehicle.
- Use the negative prompt for music. “music, score” keeps the result to effects and ambience when the model tries to add a bed.
- Keep dialogue on for talking scenes. The model detects speech and ducks the effects under it, so the lines stay clear.
- Turn dialogue off for footage with the wrong voices. Temp narration, set chatter, or a reference track you’re replacing anyway.
- Describe the space, not just the source. “Footsteps on wet concrete in an underground garage” lands better than “footsteps.”
- Generate variations and mix. Two takes of the same scene give you a layer to blend or a cleaner version of a single hit.
- Videos up to ten minutes work in one pass. Longer than that, split the cut first.