Mirelo SFX 2.0 re-sounds a 10-minute edit in one pass

Mirelo released SFX 2.0 on October 14, a video-to-sound model that replaces a finished clip's audio without regenerating the picture, in runs of up to 10 minutes in Mirelo Studio and the API, with unlimited length available on request through its Enterprise tier.

The announcement post names the problem: video models bake in music nobody asked for, re-prompting burns credits without fixing it, and muting the clip takes the dialogue with it. SFX 2.0 keeps the take and rebuilds the sound around it.

The model detects speech, keeps it, removes the rest of the original track including unwanted music, and generates foley, effects and ambience that follow the action. It holds one ambience across cuts, which matters for sequences stitched from separately generated clips. Speech preservation is on by default in the API.

What comes back is a single mixed track. SFX 1.6 could return dialogue and effects as separate stems; the 2.0 model card lists a mix stem only, and a multi-stem request is refused rather than ignored. A stem variant is coming soon, and 2.0 covers only text-to-sound and video-to-sound, so the extend and inpaint edits 1.6 offered are absent.

Hosts price it by output length. On fal, SFX 2.0 costs $0.010 a second, $0.60 a minute, through December 13, 2026, then $0.015, or $0.90 a minute. WaveSpeed charges $0.016 a second, $0.96 a minute, and bills the duration you request even when the clip is shorter. Runway lists it at 15 credits a minute. Mirelo's Creator plan is €20 a month for about 40 minutes; the free tier gives 5,000 credits, about 8 minutes, with a watermark.

Native audio is the option it competes with. On Runway's developer pricing, one second of Veo 3.1 with audio costs 40 credits and one second of Kling 3.0 Pro with audio costs 17, against 15 credits for a full minute of Mirelo. Mirelo says its own blind listening tests preferred SFX 2.0 to Veo 3.1 Fast's native audio in 57% of comparisons, to Seedance 2.5's in 47% and to Kling's in 45%, with 25 raters judging 30 videos each time.

The documentation does not say how SFX 2.0 handles a sung vocal, whether a sung line counts as speech to keep or music to strip, and one mixed output leaves no stem to check.

Availability is wide: Mirelo Studio, the Playground, the API, an MCP server, Premiere Pro and DaVinci Resolve plugins, plus Runway, fal and WaveSpeed. Roblox is coming soon.

If the rates hold, re-sounding a locked edit costs a fraction of regenerating it, and the picture you graded survives. A mixer who needs dialogue isolated on its own track waits for the stem variant or stays on 1.6.

SFX 2.0 is live in Mirelo Studio and on Runway, fal and WaveSpeed; the browser playground needs no install, and the free tier covers about 8 minutes of video.