Mirelo SFX 2.0 wants to fix one of AI video's most annoying flaws

Mirelo SFX 2.0 allows you to replace unwanted music, keep dialogue, and generate a synchronized soundtrack for videos up to 10 minutes long.

A successful video sequence can become unusable because of unrequested music, misplaced sound effects, or an ambiance that shifts between shots. Rather than regenerating the entire video, Mirelo SFX 2.0 offers to rebuild its soundtrack. The model directly analyzes the images to generate synchronized sound effects and ambiances, while preserving existing dialogue. Remove music without losing dialogue

Video generators sometimes produce a complete soundtrack, even when the user only wants dialogue and sound effects. Removing this track then means making the voices disappear as well, while generating a new version risks modifying a sequence that has already been approved.

Mirelo SFX 2.0 seeks to avoid this problem by distinguishing speech from other sound elements. The model can remove unwanted music, replace existing effects, and rebuild an ambiance around the preserved dialogue.

This preservation relies on processing the original audio track, rather than creating new lines of dialogue. It allows creators to keep the chosen performance without having to redo the entire virtual shoot. Like any audio separation process, however, it can lead to artifacts depending on the quality of the initial file. A soundtrack built from images

The other key feature of Mirelo SFX 2.0 lies in its video-to-sound operation. Unlike tools that generate effects from a text description, the model uses the video as its main source of information.

It analyzes movements, interactions, and the environment to place sound effects precisely during actions. Footsteps, impacts, object handling, or ambiances are thus generated based on what appears on screen.

The model also strives to maintain the acoustic consistency of a sequence. A car should keep a comparable sound signature across multiple shots, while the ambiance of a single location should not change arbitrarily with every cut. Up to 10 minutes of audio in a single generation

Mirelo SFX 2.0 can process an entire edit without requiring the video to be split into short sequences to generate sounds separately.

Mirelo Studio and the API accept videos up to 10 minutes long, while a duration with no predefined limit is available upon request in the Enterprise plan.

This continuity is particularly appealing to creators assembling multiple clips from different video generators. The model can process the entire edit to harmonize effects and ambiances, while leaving the choice of the final music up to the user. Listening tests against Sony, Google, and ByteDance

Mirelo also published blind comparisons conducted against other sound generation models. Against Sony Woosh, Mirelo SFX 2.0 was preferred 81% of the time, compared to 9% for its competitor and 10% with no preference.

The company also compared its results to the native soundtracks of several video generators, including Veo 3.1 Fast, Seedance 2.5, MiniMax H3 Max, Flux 3, and Kling.

Against Veo 3.1 Fast, Mirelo was preferred in 57% of cases, compared to 21% for the native audio and 22% tied. For Seedance 2.5, the proportions were 47%, 28%, and 25%.

Each comparison involved 30 videos evaluated by 25 people, based on sound quality and alignment with visible actions. These tests were organized by Mirelo and do not constitute an independent validation of its performance. Available for creators and developers

Mirelo SFX 2.0 is accessible in Mirelo Playground, Mirelo Editor, via its API and its MCP connection, as well as on partner platforms Runway and Fal.

Integrations for Adobe Premiere Pro, DaVinci Resolve, and Roblox have also been announced. A text-to-sound variation allows users to produce effects without video, while a future stem variant is expected to offer isolated sound generation for more precise work on each audio layer.

With SFX 2.0, Mirelo addresses a concrete limitation of generative video: the difficulty of modifying audio without losing the images or dialogue you want to keep. The model's value lies less in creating a soundtrack from scratch than in its ability to step in after editing, once the visual choices have already been finalized.