Creators report removing video music while preserving dialogue and effects

• Music removal took minutes while preserving the picture, dialogue and effects, according to dustinhollywood.
• A separate local workflow produces dialogue, music and effects stems using SAM Audio in ComfyUI, as QualitativAi demonstrates.
• Hollywood is keeping his workflow private while planning a Stages AI tool.

SAM Audio can use a masked object in a video to select which sound to isolate. An independent ComfyUI extension also accepts time spans marking where a target sound does or does not occur.

The client brief was precise: remove the music while leaving the video, dialogue and sound effects intact. Dustin Hollywood says he completed the audio fix in minutes, faster and cheaper than regenerating the clip.

Hollywood's clever shortcut combined a music-production technique with a video-editing process, he explained, without needing a generative model. He declined to publish the workflow, arguing that an artist's methods are part of their earning power.

In a separate demonstration, QualitativAi runs Meta's SAM Audio through a local ComfyUI graph to produce three stems:

• Dialogue
• Music
• Sound effects

An agent runs a ComfyUI workflow to separate dialogue, music and sound effects

For a repeatable interface, QualitativAi proposes asking Opus or Astra to build the workflow and a web app, with saved separation defaults. ComfyUI would run underneath without the user opening its editor.

QualitativAi reports better results than the Moises three-stem split they previously paid $30 a month for. They note that getting the local process to their liking can take some fiddling.

SAM Audio writes two files in Meta's reference implementation:

• target.wav: the isolated sound selected by the prompt.
• residual.wav: the remaining audio, where dialogue and effects would be retained when music is selected as the target.

The three-stem output is part of QualitativAi's demonstrated workflow; Meta's basic example exposes this two-output operation.

Meta also documents a quality-versus-compute trade-off: enabling automatic span prediction and reranking eight candidates can improve results while increasing latency and memory use.

Hollywood says repeated requests prompted him to turn his own method into a Stages AI tool. He hopes to have it ready by launch.

QualitativAi separately intends to open-source their workflow.

The underlying SAM Audio model is already available from Meta, including through its Segment Anything Playground. Local use has these documented requirements:

• Python 3.11 or newer.
• A CUDA-compatible GPU is recommended.
• Hugging Face checkpoint access must be requested and approved, followed by authentication to download the models.