Introduction

Mirelo AI just shipped Mirelo SFX 2.0, a video-to-sound model that rebuilds foley, ambience, and effects around the footage you already have — while keeping the dialogue and stripping unwanted music. Studio and API support clips up to 10 minutes in one pass. It’s live today on Mirelo’s Playground, Editor, API, and MCP, plus partner surfaces on fal and Runway.

What SFX 2.0 actually does

Upload a finished (or stitched) video. The default video-to-sound path:

  • Detects and preserves spoken performance
  • Removes non-dialogue audio such as baked-in music
  • Generates action-synced SFX, foley, and ambience that stay coherent across cuts
  • Delivers one soundtrack pass for the full edit (up to 10 minutes; Enterprise can go longer)

That’s the pain point for AI filmmakers and video agents: native video-model audio is often sticky music you didn’t ask for, and regenerating the whole clip just to fix sound burns credits and takes. Text-to-SFX libraries don’t see the picture, so every hit still needs hand alignment. SFX 2.0 starts from the pixels.

Blind-test claims (vendor)

Mirelo ran crowd-sourced blind listening (25 raters × 30 videos per comparison). Headline numbers from the launch post:

  • Preferred over Sony Woosh in 81% of cases (9% for Woosh, 10% tie)
  • Preferred over Kling Foley in 63% of cases
  • On cinematic AI clips, rated as good as or better than native audio from models including Seedance 2.5, Veo 3.1 Fast, MiniMax H3 Max, Flux 3, and Kling (e.g. ≥ Seedance 2.5 sound in 72% of cases when counting ties as “as good”)

Treat these as vendor-reported; still useful as a same-day quality signal for video-to-SFX.

Variants and where to run

Three product lines:

  • video-to-sound (default) — single coherent mix synced to picture, woven around existing speech
  • text-to-sound — prompt-driven SFX/ambience when picture sync isn’t required
  • stem — coming soon; one synced sound at a time for layered control

Surfaces today:

fal intro pricing (Oct 14–Dec 13, 2026): $0.010/s of output (then $0.015/s from Dec 14). Billing is per output second × samples — e.g. two 10s samples = $0.20 intro / $0.30 standard.

Why it matters for gen video

If you generate clips one shot at a time, each take often sounds like a separate world. SFX 2.0’s continuity-across-cuts pitch is aimed at agents and editors who assemble a scene first, then want one coherent post pass — dialogue kept, junk music gone, music bed left for you to choose.

Original Source

https://mirelo.ai/blog/mirelo-sfx-2-0

https://fal.ai/models/mirelo-ai/sfx2.0/video-to-audio

Conclusion

SFX 2.0 is the rare audio drop that meets AI video where it actually breaks: after the cut is locked. Try a clip with sticky music on Mirelo or fal, keep the take you like, and listen for whether the new foley lands on the action.

—Aurelia ♡