Mirelo SFX 1.6

Every video you generate is silent until you do something about it, and the something is usually a trip to a sound library or a foley session. Mirelo SFX 1.6 is the shortcut: a hosted text-to-sound-effects model. Type "rain on a tin roof, distant thunder" and it generates the audio track. This node wraps it in a ComfyUI socket that returns a native AUDIO type - so it's the missing soundtrack piece for a video workflow, and it slots in right next to the video nodes that produce the footage it scores.

Here's the honest framing: it's a niche tool for a very specific pain. If you're doing AI video, the two hardest problems after the visuals are (a) finding/creating the right sound and (b) making it loop or line up. Mirelo is aimed at (a), and it's genuinely good at it because it's a sound-effects model, not a music model - it understands "creaking door" and "crowd murmur" and "shallow breathing" in a way a text-to-music model won't.

Inputs that matter

There are no required inputs - you can run it with just a prompt. The important ones:

• positivePrompt - what you want the sound to be. Be concrete about the material and the action: "wooden chair scraping on tile" beats "chair sound."
• negativePrompt - defaults to music, and this is the setting to respect. The model's first instinct is musicality; if you want pure foley, leave the negative prompt alone so it steers away from melody.
• duration - seconds, 0 to 60. Zero means "you decide"; otherwise this is your loop-length knob.
• settings.ambiance - produce a seamlessly loopable output. This is the unsung hero for background ambience: switch it on and the clip wraps around without a seam. Pair with settings.doubleOutput to concatenate the loop with itself twice for a longer bed.
• video / audio - UUID or URL inputs, for when the effect should sync with existing media (sound a clip, or re-sound an existing track). The audioSettings socket (from the Runware Audio Settings builder) handles bitrate/channels/sample rate.
• CFGScale - default 4.5 (0–10), and steps default 25 (1–30). These are real diffusion knobs on an audio model, and unlike the image nodes they're exposed directly, not behind an enable-toggle.
• settings.repaintingStart - inpaint-style editing: repaintingStart_value sets the offset in seconds where the model regenerates from, so you can patch a bad section instead of redoing the whole track.

How it works

The node sends mirelo:[email protected] as an audioInference task through the Runware SDK. The audio output comes back as a native AUDIO socket (the pack uses soundfile/torchaudio to decode), which wires into core audio consumers or the video nodes' audio inputs - that's the whole trick for adding sound to a generated clip without any external tooling.

Install and gotchas

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

Restart; API key via Settings → "Runware API key", RUNWARE_API_KEY, or runware auth login. Paid per run, minimum top-up.

The gotcha that bites people: expecting full compositions. This is SFX, not a score. Prompts for "a sad orchestral piece" will disappoint you - it's a foley and ambience tool. Also, if you leave duration at 0 you may get a shorter clip than your video needs; set it explicitly once you know your loop length. And remember the negativePrompt: music default is load-bearing - "fix" it to empty and suddenly every effect comes back with a melody over it.