You finally get the take. The actor hits the mark, the light is right, the line lands. Then you play it with sound and a string section you never asked for swells under the dialogue, and the footsteps on a wet floor sound like slippers on carpet.
Until recently the fix was to regenerate and hope, or to strip the audio and rebuild it by hand. There is now a third option that works better than either. Keep the picture and the dialogue, and hand the clip to a video-to-sound model that rebuilds everything else to match what is on screen. The newest one worth your time is Mirelo SFX 2.0, released October 14, 2026.
Why the audio is usually the weak part
Native audio was the big shift of the spring. As we wrote in May, models now generate sound in the same pass as the picture, and that removed the silent draft from most workflows. It did not remove the sound editor's job. Video models treat audio as a side output. They improvise a score when the prompt feels dramatic, they flatten foley into generic whooshes, and they forget what the previous shot sounded like, so the rain changes character on every cut.
Regenerating to fix the sound is the expensive way out. You pay for a new video take, and you risk losing the performance you liked. At the $0.10 a second widely reported for Gemini Omni 1.1 Flash at 720p, a 10-second retake is $1.00, and the next take has its own problems.
How to add sound effects to AI video after the render
The approach that holds up is to treat the clip like location footage. The picture is locked, the dialogue is the production track, and everything else gets rebuilt in post.
That is what SFX 2.0 is built for. Mirelo says it keeps the speech and rebuilds the sound around it, removes music the video model added without being asked, and uses the footage itself as the main input. That last part is the important one. Text-to-sound tools hear only your prompt, so every door slam has to be found, trimmed and nudged onto the frame by hand. A video-to-sound model sees the door close.
Mirelo also claims the sound stays consistent when you stitch several shots into one video, which goes straight at the cross-cut drift problem. Its blind tests had listeners preferring SFX 2.0 over Sony's Woosh 81% of the time. That is a vendor number, so take it as a reason to test rather than a reason to believe.
There is also a text-to-sound mode for things the camera cannot see: a train two blocks away, a party through the wall. Use it for off-screen layers and sync those yourself.
What it costs
On fal, SFX 2.0 video-to-audio bills per second of output and per sample: $0.010 a second at the introductory rate through December 13, 2026, then $0.015. The catch is that the default request returns two samples, so a 10-second clip costs $0.20 at the intro price, not $0.10. Set the sample count to 1 once you know what you want.
At the standard rate and one sample, rebuilding the sound for a three-minute short costs about $2.70. That is cheaper than one regenerated shot on most premium video models.
A worked example
Here is how we would run it on a typical problem shot. Twelve seconds, interior, night: a waitress pushes through a diner door out of the rain, shakes off her coat and says, "Kitchen's closed, but I can do coffee."
Say the native audio gives you a clean line, a vague rain bed, no door, and a piano cue that comes in on "coffee" and stays. The fix is four steps:
- Export the scene as its own clip with the original audio attached, so the model can find and keep the dialogue.
- Run video-to-audio with two samples for the first pass. At 12 seconds that is $0.24.
- Pick the sample where the door hits on the frame it closes and the rain drops in level once she is inside. That level change is the detail a viewer feels without noticing.
- If the dialogue sits low, duck the effects bed by 3 to 4 dB under the line in your editor and leave the rest alone.
The prompt for the original shot stays the same if you regenerate later. The sound is now a separate decision.
What to do about it
Prevent the problem upstream first. Write "no music" into your video prompts unless you want it, and describe the sound you need the same way you describe a camera move: "rain on glass, muffled once the door closes, one door slam, no score."
On Promvie, background music is off by default and optional per scene, and every scene renders with dialogue and sound in the same pass. When a scene comes back with the right performance and the wrong ambience, export it and give it a sound pass instead of spending credits on a new take.
Then save video-to-sound for the shots that earned it. A good performance is the hardest thing to get out of an AI model. Do not throw one away over a door that never slammed.