Mirelo SFX 2.0: Fix the Sound on an AI Video Without Re-rendering the Picture

The fastest way to spot an AI video is to close your eyes.

The pictures have become good enough to fool people for a few seconds. The sound has not. Footsteps land a beat late. A door closes with no click. A quiet kitchen scene arrives with a swelling piano score nobody asked for, because the video model decided the moment felt emotional. Most video models now generate sound alongside the picture, which sounds like progress until you realise the two are welded together. If the music is wrong, the usual fix is to generate the whole clip again and hope the new picture is as good as the one you liked.

Mirelo SFX 2.0, released on October 14, is built on a different idea. Its launch post describes it as a model "that lets you change a video's audio without regenerating its original video." It watches the clip, keeps any dialogue, removes the other audio, and builds new sound effects, foley (the small recorded sounds of hands, fabric and feet) and room ambience timed to what happens on screen. The picture never changes.

That one design choice matters more than any score on its launch page. It means sound becomes a separate pass you can redo for pennies, and stops being a reason to throw away a good shot.

Why welded sound costs you more than it looks

Think about how a film soundtrack normally gets built. The picture gets locked first. Then a sound editor goes through it shot by shot, listing every sound the scene needs: the cup set down, the traffic outside, the moment the room goes silent. That list is called a spotting sheet, and it exists because sound is easier to get right once the picture stops moving.

AI video skipped that order. When picture and sound come out of the model together, every sound note becomes a picture gamble. Want the rain louder? Reroll. Want the score gone? Reroll, and maybe lose the expression on the actor's face that made you keep the take.

Mirelo puts the order back. You lock the picture, then you spot the sound. The model does the timing work that would otherwise mean hours of dragging library effects onto a timeline, and you keep the decision about what the scene should sound like.

What it actually does

Here is the plain version of what Mirelo's post and its API listings describe:

• Input is your video. It reads the frames and places sounds on the action. You can add a text description to steer the mood.
• Dialogue is kept by default. On WaveSpeed's listing, one of the services that sells access, the setting that preserves speech is switched on unless you turn it off.
• Other audio goes. The launch post lists removing non-dialogue audio, such as unwanted music, as a feature.
• You can say what to avoid. WaveSpeed's version takes a second, "avoid" description, so you can write "no music, no crowd noise" and mean it.
• Clips up to 10 minutes. Longer runs need Mirelo's enterprise tier.
• It lives where editors already work. The post lists Mirelo's own editor and playground, an API, plugins for Premiere Pro and DaVinci Resolve, and partners including Runway and fal. Runway added it to its assistant connector on October 15, according to its changelog, so an AI assistant can make a clip and then score it in one session.

What you get back is one finished audio track, delivered as WAV, MP3, AAC or FLAC on WaveSpeed. The picture file is not re-rendered.

What it costs

This is where the case gets strong.

WaveSpeed lists Mirelo SFX 2.0 at $0.016 per second of audio. A 10-second shot costs 16 cents. A full minute costs 96 cents. A ten-minute short costs $9.60.

One catch hides in that billing. You pay for the length you request, not the length of your video. Ask for 60 seconds on a 20-second clip and you pay for 60. Set the length to match the clip every time.

Mirelo's own pricing page offers a free tier of 5,000 credits a month, which it estimates at about eight minutes of SFX 2.0. Free exports carry a watermark and free projects are deleted after a week, so the free tier is for testing, not delivery. The monthly prices for paid plans did not load when I checked.

Compare that to the alternative. If a clip's picture is right and its sound is wrong, rerolling means paying full video prices again, waiting again, and gambling the picture. A sound pass costs a fraction of that and risks nothing you have already approved.

Put this into practice

Pick one AI clip you already like but never used because the sound was off. Then run this routine.

1. Lock the picture first. Finish the cut, the trims and the colour before you touch sound. Every sound pass is timed to the frames you give it, so changing the edit afterward means paying for another pass.

2. Write a spotting sheet, not a vibe. Watch the clip once with the sound off and write down every sound you expect to hear, in order. "Rain on a car roof, wipers twice, keys dropped on the seat, engine turns over on the second try." Turn that list into your prompt. Specific nouns tell the model what to place. Mood words like "cinematic" or "tense" tell it almost nothing it can time.

3. Say what you do not want. Use the avoid field for the habits you are trying to escape: "music, score, crowd chatter, whooshes." Video models love to add music. Name it so it stays out.

4. Match the length exactly. Set the requested duration to the clip length. You pay for every requested second.

5. Generate three, keep one. The same request can come back different each time. Run three passes, listen on headphones and on a phone speaker, and pick the take where the key moment lands. When you find one you like, note its seed number (the value that makes a result repeatable) so you can get back to it.

6. Mix it like any other track. Drop the audio under the picture in your editor and treat it as a raw sound bed. Pull it down under dialogue, add a real music cue if the scene needs one, and cut any effect that draws attention to itself.

7. Check it against the original. Listen to the old and new tracks back to back, with the dialogue soloed if you can. If a line got clipped or a word went missing, that is a reason to rerun, not to ship.

If you work in Premiere Pro or Resolve, the plugins let you do this without leaving the timeline. If you work with an assistant inside Runway, you can ask for the sound pass in the same conversation that made the clip.

The honest limits

You get one mixed track. There are no separate layers for rain, footsteps and room tone yet. Mirelo marks a stem version (separate layers you can mix individually) as "coming soon." Until then, if the rain is too loud, you rerun the pass rather than turning one fader down. That is the biggest gap for anyone who mixes for a living.

The quality scores are Mirelo's own. The launch post reports blind listening tests where its sound beat Sony Woosh 81% to 9% and beat MiniMax H3 Max's built-in audio 49% to 30%, with the rest saying no preference. Those come from 25 crowd-sourced listeners judging 30 videos, picked and run by the company selling the product. They are a reason to test, not a reason to trust. Notice also that "beats the video model's own sound" still means a lot of listeners had no preference or preferred the original.

Dialogue protection is a setting, not a promise. Keeping speech is the default, but the model decides what counts as speech. Whispered lines, singing, and speech buried under noise are the places to listen hardest.

Commercial rights were not clear when I checked. Mirelo's terms page did not load in a readable form, and its pricing page does not spell out usage rights for each plan. The free tier's watermark rules it out for client work anyway. Read the terms on whichever service you pay before you deliver anything.

It cannot fix a picture problem. If the actor's feet never touch the ground, perfect footstep timing will only make that more obvious. Sound follows picture. A broken shot stays broken.

Where this leaves you

Sound has been the part of AI video that people accept because fixing it seemed to cost a whole new render. Mirelo SFX 2.0 changes that math. A sound pass now costs about a cent and a half per second, leaves the picture alone, and plugs into the editors you already use.

The larger shift is in how you should plan. Generate the picture until it is right, then stop paying the video model for sound you will replace. Spend your attention where a sound editor would: the list of what the scene should actually sound like.

Take the clip you shelved because it sounded fake. Write its spotting sheet tonight, run it through the free tier tomorrow, and listen with your eyes closed. If you cannot tell, you just rescued a shot. If you can, you learned exactly where the tool stops, for less than the price of a coffee.