Runway scores video without touching the dialogue

Runway has hooked Mirelo SFX 2.0 up to its platform — a model that builds a soundtrack from the image: it reads the action, timing and acoustics of the scene, places footsteps, impacts, foley and ambience against what happens on screen, and preserves the speech in the clip while doing so. Mirelo released version 2.0 on October 14, 2026; Runway announced its arrival on its own platform two days later.

For the generative video pipeline this closes a weak link. The picture is assembled with models, but sound remains handwork: the editor takes the clip into Premiere Pro or DaVinci Resolve and manually hunts for, trims and lines up library effects. Mirelo proposes doing that pass automatically — on top of an already finished edit, without regenerating the video.

Speech turned out to be harder than the effects themselves

The main problem in automatic sound design is not synthesizing footstep noise but avoiding trampling speech that has already been recorded: automatic SFX tools usually lay the generated layer over the dialogue. Mirelo SFX 2.0 detects speech in the clip and preserves it in the final track, clearing out the rest of the non-dialogue audio — including music the video model added on its own — and builds the new effects around the lines, handling ducking as well.

Speech preservation is on by default for video-to-SFX requests; it can be turned off if only effects are needed. There is no separate "remove music" toggle — removal of non-dialogue audio is built into the model's own logic.

What the model can do and where its limits are

The model has shipped in three variants:

• video-to-sound — the main mode: a single coherent sound layer, synchronized to the video and built around the existing speech.
• text-to-sound — for when synchronization to picture is not needed (ambience, for example) or when off-screen sound has to be added and lined up by hand afterwards.
• stem — generating one synchronized sound at a time, to assemble the mix layer by layer; the mode is announced as coming.

Through Mirelo Studio and the API the ceiling is 10 minutes of video; the company states that the sound holds together across that entire length and does not fall apart at cut boundaries, so a finished edit does not have to be chopped into short fragments. For longer material an enterprise plan is offered on request.

How much better is this than what already exists

Mirelo has published the results of a blind comparison: 25 recruited listeners per pairing, 30 clips, a choice between two tracks on sound quality and fit to what is happening on screen, with a "tie" option. Listeners were recruited on a crowdsourcing platform with pre-screening for accuracy on audio tests. By those figures, SFX 2.0 was preferred over the Sony Woosh track in 81% of cases, over Kling Foley in 63%, over Sonilo in 46%. Against the sound video models generate themselves the gap is smaller: Veo 3.1 Fast — 57%, MiniMax H3 Max — 49%, Flux 3 — 48%, Seedance 2.5 — 47%, Kling — 45%. The test was run by the vendor itself and there are no independent measurements so far, so the numbers are worth treating as the company's claim rather than an outside measurement.

Where to get it

Besides Runway and fal, Mirelo SFX 2.0 is available in the company's own playground and editor, through an API and MCP, and as plugins for Adobe Premiere Pro and DaVinci Resolve; a version for Roblox has been announced as coming. The company promises to connect the remaining platforms later.

Why Runway is taking someone else's model for sound

A platform with its own audio generation is adding a third-party one. Runway has its own tools for sound and video models that output audio natively, but all of them start from a prompt or from their own video generation, whereas Mirelo works from a finished frame — including one that came from someone else's model or from a camera. For a platform that methodically plugs in partner models, this is a logical addition: sound stops being a by-product of the video generator and becomes a separate layer that can be redone without touching the picture.

A broader shift shows through behind this. While video models were learning to output sound together with the picture, part of the industry bet on the opposite — on specialized models for the audio layer that work after the edit. The logic from both sides converges on one point: splitting a generative clip into controllable layers instead of one indivisible output. Mirelo itself is explicitly aiming at this market — its API is sold to teams building video agents and video-generative platforms.

The practical upshot: the routine part of scoring — rough foley and ambience under a finished edit — becomes a matter of a single pass. It most likely presses hardest on stock effects libraries and contractors on simple jobs. The winners are editors and small teams who previously had to choose between silence and someone else's music. Sound-engineering work remains, however: the model hands over a summed layer that still has to be fitted into the mix, and layered output is so far only on the vendor's roadmap.

Which part of Mirelo SFX 2.0 will you put to work first?

• Removing music without losing speech
• Foley to picture