Mirelo, an AI audio company, has launched Mirelo SFX 2.0, a “video-to-sound” model that generates new sound for an existing video without having to regenerate the video itself.
In a blog post on 14 October 2026, Mirelo said the model “preserves the dialogues, removes unwanted music, and generates sound effects” that are synced to the action on screen. It also generates foley (everyday sounds such as footsteps or doors) and ambience, and Mirelo says the sound stays consistent across cuts for videos up to 10 minutes long.
Why it matters for AI video
Many AI video generators create audio along with the picture. Mirelo says that means you cannot change the sound without regenerating the whole clip, which is slow and costly. Text-to-sound tools do not see the video, so each effect has to be lined up by hand. SFX 2.0 uses the footage itself as its main input, so it can match sounds to what happens on screen.
Mirelo’s own tests
Mirelo says it ran blind listening tests in which people heard two soundtracks for the same video and picked the one they preferred. For each comparison, 25 raters recruited on a crowd-sourcing platform judged 30 videos. These are Mirelo’s own tests, not independent ones.
In those tests, according to Mirelo, listeners preferred SFX 2.0 over Sony’s Woosh model in 81% of cases (9% preferred Woosh; 10% had no preference). Against other video-to-sound models, Kling Foley and Sonilo, its preference rates were 63% and 46%. Compared with the sound that AI video models generate themselves, it was preferred in 57% of cases against Veo 3.1 Fast and in 47% against Seedance 2.5, with the rest either preferring the original or rating them equal.
Where to get it, and the price
Mirelo says SFX 2.0 is available in its own editor and playground, through its API and MCP server, and on Runway and fal. Plugins for Adobe Premiere Pro, DaVinci Resolve and Roblox are marked as coming soon.
On fal, a platform for running AI models, the introductory price is $0.010 per second of output from 14 October to 13 December 2026, rising to the standard $0.015 per second from 14 December, according to fal’s model page.
What this means
- Sound becomes editable separately. Creators can change audio without regenerating a video.
- Dialogue stays. The model keeps speech while replacing music and effects.
- The comparisons are the company’s own. No independent tests have been published.
Sources
- Mirelo, “Mirelo SFX 2.0: change the sound, keep the video”, 14 October 2026. https://mirelo.ai/blog/mirelo-sfx-2-0
- fal, Mirelo SFX 2.0 video-to-video model page. https://fal.ai/models/mirelo-ai/sfx2.0/video-to-video
