What Was Announced on October 16, 2026
At 19:56 UTC on October 16, 2026 the @runwayml account on X posted: "Mirelo SFX 2.0 is now on Runway, generating full soundscapes from your video while keeping dialogue intact." Mirelo itself published its release material two days earlier, on October 14, 2026, on its blog.
Two claims sit in that announcement: sound is generated from video, and dialogue stays untouched. The second one matters most in practice, because most tools that touch audio break speech first.
What Mirelo SFX 2.0 Does
It is a video-to-sound model. Video goes in, and a soundscape comes out: effects, foley and ambience synced to the action on screen. Per Mirelo's description, the model:
That last point rarely gets mentioned and is the most practical one. When a clip is assembled from several generations, sound normally collapses at the seams: one scene sits in silence, the next carries foreign ambience. Here the sound is generated across the whole timeline rather than per fragment.
- preserves dialogue exactly as recorded;
- removes unwanted music;
- generates SFX, foley and ambience in sync with the action;
- stays coherent across cuts, so the sound does not fall apart at edit points.
Three Modes and the Length Limit
Length: up to 10 minutes in Studio and API, with unlimited duration available on request on the Enterprise tier. Ten minutes is not about vertical clips or ad inserts; it is long-form territory such as reviews, training material and webinar recordings, where sound has to be built on top of an already edited video.
- video-to-sound, the primary mode, sound from video;
- text-to-sound, sound from a written description;
- stem, announced as coming soon.
Where It Is Available
There are several platforms, and that is a deliberate strategy: Mirelo is not locking sound inside its own interface.
For production, two routes are the practical ones: the API if you already run a pipeline, and the editing plugin once it ships. In between sits Runway, where the tool lives next to video generation.
| Platform | Status |
|---|---|
| Mirelo Editor | available |
| Mirelo Playground | available |
| Mirelo API | available |
| Mirelo MCP | available |
| Runway | available since October 16, 2026 |
| Fal | available |
| Premiere Pro, DaVinci Resolve, Roblox | plugins, coming soon |
Blind Tests: What the Vendor Claims
Mirelo reports blind test results, and these are vendor data: 25 raters, 30 videos.
There is a separate block comparing against the native audio that video models generate along with the picture:
Read the numbers exactly as presented: vendor-run tests, not independent evaluation. A 45 to 57 percent share of votes for Mirelo is a consistent advantage, not a rout: in 21 to 30 percent of cases raters chose the audio the video model produced itself, and the rest saw no difference.
| Comparison | Preferred Mirelo | Preferred competitor |
|---|---|---|
| Against Sony Woosh | 81% | 9% |
| Against Kling Foley | 63% | 18% |
| Against Sonilo | 46% | 33% |
Why Business Video Needs Separate Sound
The first reason is the big one: models bake music into the video. Generators often add audio together with the picture, and when that music does not fit the brand or the platform, regenerating the sound is cheaper than regenerating the whole clip.
The second reason is assembly from several generations. Making videos with AI rarely yields one continuous scene: a clip is built from pieces, and the sound has to be continuous. A model that works across the timeline removes manual audio stitching.
The third reason is platform versions. The same video may go to different channels with different music and ambience, and that is cheaper than rebuilding the picture.
Sound and Video Versions
There is a nuance worth stating plainly. In versioning work, sound is one of the layers that separates one version from another: different music, different ambience, different rhythm in the opening seconds. For a series of mass vertical videos this is ordinary versioning work, on par with the cover frame and the opening shot.
The two jobs should not be confused, though. Versioning works with variants of the same material so a series does not read as one file; it is not about erasing origin markers. We covered how detection differs from versioning separately, and different sound belongs to versioning, not to dodging checks.
What to Do This Week
For market context on models and price per second we keep a 2026 AI video model guide. Our own work is priced in dollars: manual videos with complex staging start at $190, and batch formats run at $2.40 per 1000 views.
- Take a finished clip with music that does not fit and run it through video-to-sound.
- Check that dialogue survives without artifacts: it is a stated behavior, but verify it on your own material.
- Build a test scene from three cuts and judge sound coherence at the seams.
- Work out which is cheaper: regenerating the whole clip or replacing only the audio.
- Decide whether you need the API or an editing plugin will do.