The locked-track problem
Native audio solved one problem in AI video and created another. Models such as MiniMax H3, Veo and Kling now generate sound with the picture, which is useful until the sound is wrong. A clip can arrive with music nobody asked for, effects that sit loosely on the action, or ambience that changes when the next clip starts. Fixing it the native way means prompting again. That costs credits, takes time, and risks losing a take the creator already liked. Muting is not a fix either, because the dialogue goes with the music.
SFX 2.0 is built for that specific dead end. The picture is treated as finished. The model reads the video, detects the dialogue, strips the non-dialogue audio, and builds a new layer of effects, foley and ambience around the speech, timed to the visible action and shaped to sit under it. One pass can cover a stitched edit, not only a single clip, up to 10 minutes in Studio and the API.
One detail in Mirelo's own FAQ deserves attention. Asked how SFX 2.0 keeps the dialogue, Mirelo says it detects the spoken performance in the clip and regenerates it together with the new effects and ambience woven around it. The words and the chosen take survive; the performance is re-rendered as part of the new soundtrack, not passed through untouched. Readers who need the original recording bit-for-bit should know that distinction before they rely on the word "preserves".
Why one pass across cuts matters
Most AI video is still assembled, not generated in one take. A product demo, a short ad or a scene is several generations edited together. Native audio makes each generation sound like its own room. Ambience shifts, effects change character, and the join is audible even when the picture cuts cleanly.
A post-production pass reverses the order of work. Generate and edit the picture first, lock the cut, then run the sound once over the whole sequence. Mirelo is also selling that pass to other platforms and to video agents through its API, which is a different bet from selling another generator. If agents are going to stitch clips without a human in the loop, something still has to make the result sound like one piece of work. SFX 2.0 is positioned as that step, and its presence on Runway and fal puts it inside workflows creators already use rather than in a separate destination.
There are limits in the launch itself. The stem variant, which would generate one sound at a time for editors who want layer-by-layer control, is listed as coming soon, so the current offer is a coherent full pass rather than full sound-design control. Anything past 10 minutes is an Enterprise conversation, not a self-serve feature. Pricing needs care too. The free tier on Mirelo's pricing page, checked on 15 October, includes 5,000 credits a month shared across Mirelo's products — usage varies by model and by Studio versus API, extra credits run €1.20 per 1,000, and the page publishes no per-generation rate for SFX 2.0 in Studio; it also carries watermarked exports and deletes projects after a week. On fal the price is explicit instead: $0.010 per generated second until 13 December 2026, then $0.015 per second — fal's hosted inference price, not Mirelo's Studio credit rate.
Read the preference numbers as vendor numbers
The launch post reports blind listening tests in which SFX 2.0 was preferred over Sony's Woosh model in 81% of cases; the full split in Mirelo's chart is 81% for SFX 2.0, 9% for Woosh and 10% with no preference. Against native audio, SFX 2.0 was rated as good as or better than the native sound of Seedance 2.5 in 72% of cases. Mirelo says listeners also preferred it on average over the native audio of every video model it tested, a list that includes Veo 3.1 Fast, MiniMax H3 Max, Flux 3 and Kling.
Those are the maker's own tests. Mirelo does state the method: 25 raters, 30 videos per comparison, recruited on a crowdsourcing platform and pre-screened for audio listening tests. That is more method than many launch posts offer, and it is still a small sample chosen, run and reported by the company selling the model. The two headline figures are also not the same measurement. One is a straight preference share against a rival sound model. The other counts ties as a pass against a video model's native audio. Read them as a reason to test the tool on your own clip, not as a settled ranking of sound models.
The practical test
The useful question is narrower than the launch claim. Take a generated clip whose picture you would keep and whose music you would not, run it through the Playground or a partner platform, and play the original and the processed version side by side. Listen for three things: whether the dialogue survives intact, whether the effects land on the action rather than near it, and whether a second clip in the same setting sounds like the same place. SFX 2.0 succeeds or fails on that comparison, and it is a comparison you can run without spending another generation credit on the picture.
Sources
- [1] Mirelo — SFX 2.0 launch post (14 October 2026; capabilities, 10-minute limit, variants, availability, FAQ on dialogue, blind-test method and results)Read source
- [2] Mirelo — pricing page (checked 15 October 2026; free-tier credits and terms, per-1,000-credit top-up, Enterprise custom pricing)Read source
- [3] fal — Mirelo SFX 2.0 video-to-audio API page (checked 15 October 2026; per-second pricing and the introductory window to 13 December 2026)Read source
- [4] HeadsUpAI — daily AI news digest (15 October 2026; independent roundup listing the SFX 2.0 launch and its availability via Mirelo, Runway and fal)Read source