Mirelo SFX 1.5
Type 'door creak' and get actual door-creak audio
- audioSettings
- audio
Sound effects are the last thing people automate, and it shows - most generated videos ship with silent scenes. Mirelo SFX 1.5 fixes that in either of two ways: type "a match being struck" and get an audio file, or feed it a video and get sound effects synced to what's happening on screen. Both run on Runware's cloud, so no audio model, no GPU, no VAE-of-sound to install.
How it works
An audioInference node. Two generation modes:
- Text-to-SFX: leave
videoempty, write apositivePromptlike "heavy rain on a tin roof," and pick aduration(1–10 seconds, default 10). TheoutputFormatchoices include MP3/WAV/FLAC/OGG - audio only. - Video-to-SFX (foley): give it a
video(a URL or Runware-hosted UUID, max 10 seconds - a string input, not a VIDEO tensor) and it produces sound that matches the action, withsettings.startOffset/settings.startOffset_valueletting you jump to a point in the clip. With a video input,outputFormatflips to video containers (MP4/MOV/WEBM) - you get the clip with audio laid in.
The audioSettings socket takes a Runware Audio Settings builder node (grab it from the Runware/Params menu) if you want to control things like the speaker/voice side of generation. Otherwise the widget knobs you'll actually touch: seed for reproducibility, steps (default 25, 5–30) for quality-vs-speed, duration for length, and numberResults for variations.
Output is an audio socket (AUDIO type, thanks to the soundfile dependency the pack installs). Wire it into a node that consumes AUDIO - a VHS audio preview/save will do - or let a video node handle it. The pack degrades gracefully to a file path if the native AUDIO type isn't available.
Install
Part of ComfyUI-Runware:
cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt
Restart (or ComfyUI Manager → search "Runware"). The requirements pull in soundfile, which is what gives audio its native type - if you installed the pack a while ago and audio nodes act weird, a pip install -r re-run is the fix. API key: Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.
Gotchas
Each run is a paid cloud inference - the title bar shows cost. SFX models are getting good, but don't expect film-quality foley on the first take; numberResults up to 4 gives you options to pick from. And remember the video input is a string (URL/UUID), so a local clip needs hosting or a Runware upload before Mirelo can watch it.
Inputs (12)
| Name | Type | Default | Description |
|---|---|---|---|
| audioSettingsopt | RUNWARE_AUDIOSETTINGS | — | |
| videoopt | STRING | Video input (UUID or URL). Max 10 seconds. | |
| positivePromptopt | STRING | Text prompt describing elements to include in the generated output. | |
| seedopt | INT | 00–2147483647 | Random seed for reproducible generation. When not provided, a random seed is generated in the unsigned 32-bit range. |
| stepsopt | INT | 255–30 | Total number of denoising steps. Higher values generally produce more detailed results but take longer. |
| durationopt | FLOAT | 10.001–10 | Length of the generated audio track in seconds. |
| numberResultsopt | INT | 11–4 | Number of results to generate. Each result uses a different seed, producing variations of the same parameters. |
| settings.startOffsetopt | BOOLEAN | false | Enable to set settings.startOffset. Off uses the model's default. |
| settings.startOffset_valueopt | INT | 00–10 | Start time offset in seconds within the input video. |
| ttlopt | BOOLEAN | false | Enable to set ttl. Off uses the model's default. |
| ttl_valueopt | INT | 60 | Time-to-live (TTL) in seconds for generated content. Only applies when `outputType` is `URL`. |
| outputFormatopt | COMBO | MP3 | File format for the generated output. Prompt-based generation produces audio (MP3, WAV, FLAC, OGG). Providing a video input produces video (MP4, MOV, WEBM). |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| audio | AUDIO | — |