AI video models now generate sound along with the picture, and that creates a new problem: if the music is wrong or the ambience clashes, you cannot fix the audio without regenerating the clip. On October 14, 2026, Mirelo launched SFX 2.0 to solve exactly that. Its announcement describes a video-to-sound model that "lets you change a video's audio without regenerating its original video": it keeps the dialogue, removes unwanted music, and generates effects, foley and ambience synced to the action, for videos up to 10 minutes.
The claims are specific and testable, and they come with a blind-test table. This post summarizes the launch, shows the numbers and explains how much weight they deserve, and says when it fits a pipeline. The product started life in August 2025 as Mirelo SFX on fal, a model you chained after a video generator to add sound to silent clips, limited to 10-second inputs. Version 2.0 moves it toward finishing the sound of a whole edit.
From 10-second clips to 10-minute edits
The history explains why this release matters. Mirelo's blog lists SFX 1.5 and Mirelo Studio in October 2025, a $41 million seed round co-led by Index Ventures and Andreessen Horowitz in December 2025, a Runware launch in May 2026, and SFX 1.6 later that month, which added longer generations, extension, seamless loops and inpainting. In July the company shipped an MCP server so assistants could generate sound effects, and in September it published guides on adding sound to video without losing dialogue. The jump from a 10-second input limit at launch to 10 minutes in one pass is the headline change: it removes the chunk-and-reassemble step that made earlier tools awkward for finished edits. It also means the model must keep ambience stable over far longer stretches, which is a harder consistency problem than syncing a single splash to a single frame.
At a glance
| Question | Answer (per Mirelo) |
|---|---|
| What does it do? | Keeps dialogue, removes other audio such as music, generates synced effects, foley and ambience |
| Maximum length | 10 minutes in Studio and API; unlimited on request for Enterprise |
| Variants | video-to-sound (default), text-to-sound, stem (coming soon) |
| Where | Editor, Playground, API, MCP, Premiere Pro and DaVinci Resolve plugins; Roblox plugin coming soon; Runway and fal |
| Evidence | Vendor blind tests, 25 raters per comparison, 30 videos |
| Price | Not stated in the announcement |
Why this matters now
Video models such as Seedance 2.5, Veo 3.1 Fast, MiniMax H3 Max and Kling increasingly produce audio natively. Mirelo points out the downside: when the model co-generates sound and picture, you cannot edit one without regenerating the other, which is "time consuming, expensive and frustrating." Muting the clip loses the dialogue; prompting again burns credits and may lose a take you liked. Sound libraries and text-to-sound tools do not look at the video, so every effect has to be found, trimmed and aligned by hand.
That is also a problem for video agents that stitch many generated clips together. Each clip arrives with its own ambience, and music baked into one shot clashes with the next. SFX 2.0 targets this workflow: finish the picture first, then run a single sound pass. For background on the video models involved, see our coverage of fal's H3 Max, Higgsfield's Seedance 2.5 offer and Kling 4.0.
What SFX 2.0 does to a video
Per the announcement, you upload a video and the model will:
- Detect and preserve the dialogue.
- Remove existing non-dialogue audio, such as unwanted music.
- Generate effects, foley and ambience that are synced to the action, fit the speech (including ducking) and stay coherent across cuts for up to 10 minutes.
- Create one sound pass for a full-length edit, rather than processing short chunks.
Mirelo says it reads the visible action, timing and environment, so a sound should differ between a park and a cathedral. It frames the output as "post-production quality." Whether that holds for your footage is a listening question; the sound is the product and the claims are subjective.
Coherence across cuts
When clips are generated separately and stitched, the ambience tends to shift between shots of the same place. SFX 2.0 aims to generate continuity across clips so several shots of one environment sound connected. This is the feature most relevant to agent-built films and ads, where the stitching happens without a human editor.
The blind-test numbers
Mirelo published two tables. Raters heard two soundtracks for the same video and picked the one they preferred for audio quality and match with on-screen events, with a "tie" option. Each comparison used 25 raters and 30 videos, recruited on a crowd-sourcing platform and pre-screened for accuracy in audio listening tests.
Against other video-to-sound models
| Competitor | SFX 2.0 preferred | Competitor preferred | No preference |
|---|---|---|---|
| Sony Woosh | 81% | 9% | 10% |
| Kling Foley | 63% | 18% | 19% |
| Sonilo | 46% | 33% | 21% |
Against the native audio of video models
| Video model | SFX 2.0 preferred | Native audio preferred | No preference |
|---|---|---|---|
| Veo 3.1 Fast | 57% | 21% | 22% |
| MiniMax H3 Max | 49% | 30% | 21% |
| Flux 3 | 48% | 23% | 29% |
| Seedance 2.5 | 47% | 28% | 25% |
| Kling | 45% | 30% | 25% |
Mirelo's own summary adds that SFX 2.0 was "rated as good as or better than Seedance 2.5's sound in 72% of cases", which is the 47% preferred plus the 25% tie.
How much to trust them
- They are vendor tests. Mirelo ran and published them; no independent lab is cited.
- Sample sizes are modest. 30 videos per comparison is small, so differences of a few points, such as 45% vs 47%, are not meaningfully different. The big gaps, like 81% against Woosh, are more informative.
- The clips matter. Mirelo describes cinematic AI video clips; results on live-action footage, music-heavy scenes or dialogue-dense scenes may differ.
- Preference is not accuracy. Listeners judged quality and match, not whether each sound is plausible for a specific scene.
- Native audio may include speech. The comparison with native audio of a video model includes whatever speech and effects it generated. Ask whether the test used the dialogue from the same clip.
The practical read: SFX 2.0 looks competitive with or better than native audio for most of the models tested, and well ahead of two competing video-to-SFX models. Treat it as a strong claim to verify on your own footage, not as settled.
The three variants
| Variant | Use it when | Status |
|---|---|---|
| video-to-sound | You want the highest-quality single coherent output synced to your video and woven around speech | Default, live |
| text-to-sound | Sync is not needed, such as ambience, or for off-screen sounds you will place by hand | Live |
| stem | You need one sound at a time, synced to video, to add a missing layer or build the mix layer by layer | Coming soon |
If you already have a music track you love, the stem variant, once available, is the likelier route; the default model removes non-dialogue audio and builds its own sound pass, so you leave music to your editor or agent.
Where it runs
SFX 2.0 is live in Mirelo Playground, Mirelo Editor, the Mirelo API and Mirelo MCP, and as plugins for Adobe Premiere Pro and DaVinci Resolve, with a Roblox plugin coming soon. Mirelo also lists Runway and fal as partners, with others coming. For teams building products, the API is aimed at video agents, video-generation platforms and software for studios and advertisers. If your agent stack already uses MCP, our MCP explainer shows how tools like this plug in. Mirelo's earlier MCP launch lets assistants such as Claude, ChatGPT and Cursor generate sound effects.
Questions people are asking
What does it cost? The announcement does not state pricing. Check the Mirelo site and partner platforms such as fal and Runway, whose per-second pricing may differ.
Is dialogue regenerated or kept as is? Mirelo's FAQ says SFX 2.0 detects the spoken performance and "regenerates it together with new effects and ambience woven around it". The headline says it preserves dialogue; the FAQ wording suggests it passes through a regeneration step. Listen for changes in timbre or timing on important takes, especially if the voice is a cloned or branded one.
What if there is no speech? The model still generates effects and ambience from the picture; the dialogue handling simply does not apply.
Does it handle licensed music? It removes existing non-dialogue audio, so if your cut contains licensed music, you would need to re-add it afterwards.
Is it safe to automate? For agent pipelines, add a review step: automated sound can introduce an effect that is wrong for a scene, and a silent mistake is easy to ship.
Bottom line
Mirelo SFX 2.0 is aimed at a real gap, fixing the sound of AI-generated video after the picture is locked, and it offers sensible controls: keep speech, drop music, one pass up to 10 minutes. The vendor's blind tests are encouraging but small. Try it on a clip whose audio you would otherwise replace, listen with and without the dialogue, and compare against the native audio from your video model before building it into a workflow.
Details reflect Mirelo's October 14, 2026 announcement; availability, pricing and variants may change.