Mirelo SFX 2.0 is a video-to-sound model for creators and video agents working with AI generated footage. It takes an existing clip, detects and preserves the dialogue, strips out the baked-in music that video models tend to add unprompted, and generates sound effects, foley, and ambience that stay synced to the action and coherent across cuts. Mirelo Studio and the API support videos up to 10 minutes, with unlimited duration available through the Enterprise tier.
The launch targets a specific pain point in current AI video workflows. When a video model co-generates audio along with the picture, you cannot change the audio without regenerating the whole clip, and stock libraries and text-to-sound tools do not help much because they do not see the video, so every effect has to be found, trimmed, and lined up by hand. By reading the footage directly, Mirelo aims to match the scene's acoustics, duck around speech, and deliver one sound pass for a full edit, which also matters for agent pipelines that stitch clips together without a human in the loop. In blind listening tests, Mirelo reports that SFX 2.0 was preferred over Sony's Woosh in 81% of cases, and on cinematic AI clips it was rated as good as or better than the native audio of models including Seedance 2.5, Veo 3.1 Fast, MiniMax H3 Max, Flux 3, and Kling.
Mirelo was founded by former musicians Carl Johann Simon-Gabriel and Florian Wenzel, who met as AI researchers at Amazon , and the Berlin-based company recently raised $41 million in seed funding led by Index Ventures and a16z, with backing from Mistral and Hugging Face executives . The timing of SFX 2.0 lands as generative video tools proliferate and sound design remains the weakest link in most AI pipelines, which is the gap this release is explicitly built to close.
Comments (10)
Sign in to join the discussion.
N
Nadia Ravasz4h ago
The 'cuts the music you never prompted' line is doing a lot of heavy lifting in that tweet. Half my Veo clips come with a cursed lounge piano I did not ask for, so yes please.
K
Kweku Mensah4h ago
Cool demo but what does the actual delivery look like? Rendered MP4 download, stems, or just a preview url that expires in 24 hours?
P
Priya S.4h ago
Before I pipe client footage through this, where does the video sit while you process it and how long do you retain it? Asking for every legal team on earth.
T
tomascuts4h ago
That launch video edit is tight. The cut from silent b-roll to full sound design at 0:08 is the entire pitch and they knew it.
R
Rhea Kothari3h ago
Already texted two portfolio founders doing AI film pipelines about this. Audio has been the ugly stepchild of generative video for way too long.
A
Amara Oluwole3h ago
First 3 seconds of that launch video are muted on purpose and that's the whole hook. Whoever edited it understands retention better than most growth people I know.
L
Linh Vo4h ago
Docs or it didn't ship. Where's the endpoint, what's the rate limit per minute of video, and do you webhook back when a long clip finishes rendering?
B
Boris Yemelin3h ago
Suspicious that every sample clip is 8 seconds. I want to see a 3 minute scene with dialogue overlap before I believe the 'consistent across any length' claim.
D
Danilo Marchetti3h ago
We tried something like this on an internal video team in 2019 and the music stripping was the hardest part. Curious what model you're using to separate score from diegetic sound without eating the dialogue.
Z
zenkoan3h ago
You can fake the picture. You cannot fake the room tone.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Mirelo
AI videos finally sound real
AI videos finally sound real
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Mirelo SFX 2.0 is a video-to-sound model for creators and video agents working with AI generated footage. It takes an existing clip, detects and preserves the dialogue, strips out the baked-in music that video models tend to add unprompted, and generates sound effects, foley, and ambience that stay synced to the action and coherent across cuts. Mirelo Studio and the API support videos up to 10 minutes, with unlimited duration available through the Enterprise tier. The launch targets a specific pain point in current AI video workflows. When a video model co-generates audio along with the picture, you cannot change the audio without regenerating the whole clip, and stock libraries and text-to-sound tools do not help much because they do not see the video, so every effect has to be found, trimmed, and lined up by hand. By reading the footage directly, Mirelo aims to match the scene's acoustics, duck around speech, and deliver one sound pass for a full edit, which also matters for agent pipelines that stitch clips together without a human in the loop. In blind listening tests, Mirelo reports that SFX 2.0 was preferred over Sony's Woosh in 81% of cases, and on cinematic AI clips it was rated as good as or better than the native audio of models including Seedance 2.5, Veo 3.1 Fast, MiniMax H3 Max, Flux 3, and Kling. Mirelo was founded by former musicians Carl Johann Simon-Gabriel and Florian Wenzel, who met as AI researchers at Amazon , and the Berlin-based company recently raised $41 million in seed funding led by Index Ventures and a16z, with backing from Mistral and Hugging Face executives . The timing of SFX 2.0 lands as generative video tools proliferate and sound design remains the weakest link in most AI pipelines, which is the gap this release is explicitly built to close.
Comments (10)
The 'cuts the music you never prompted' line is doing a lot of heavy lifting in that tweet. Half my Veo clips come with a cursed lounge piano I did not ask for, so yes please.
Cool demo but what does the actual delivery look like? Rendered MP4 download, stems, or just a preview url that expires in 24 hours?
Before I pipe client footage through this, where does the video sit while you process it and how long do you retain it? Asking for every legal team on earth.
That launch video edit is tight. The cut from silent b-roll to full sound design at 0:08 is the entire pitch and they knew it.
Already texted two portfolio founders doing AI film pipelines about this. Audio has been the ugly stepchild of generative video for way too long.
First 3 seconds of that launch video are muted on purpose and that's the whole hook. Whoever edited it understands retention better than most growth people I know.
Docs or it didn't ship. Where's the endpoint, what's the rate limit per minute of video, and do you webhook back when a long clip finishes rendering?
Suspicious that every sample clip is 8 seconds. I want to see a 3 minute scene with dialogue overlap before I believe the 'consistent across any length' claim.
We tried something like this on an internal video team in 2019 and the music stripping was the hardest part. Curious what model you're using to separate score from diegetic sound without eating the dialogue.
You can fake the picture. You cannot fake the room tone.