Replicate declare-lab/tango

Tango is a text-to-audio model: you type a description of a sound - "quiet speech and then an airplane flying away," "rain on a tin roof with distant thunder" - and it generates a matching audio clip. Not music with structure and melody (that's more MAGNeT's lane), but ambient sound and sound effects, the kind of thing you'd layer under a video. This node lets you generate that clip from inside a ComfyUI graph instead of a separate tool.

It's part of ComfyUI-Replicate, so the model runs on Replicate's cloud, not on your GPU. You send a text prompt, their hardware runs Tango, you get audio back.

How it works

Tango is a latent diffusion model that denoises in an audio latent space, conditioned on your text via a language model. In practice you don't need to care about the internals - the node takes a prompt and settings, calls Replicate, and returns an AUDIO output. Nothing downloads locally; you pay per run and there's cloud latency.

Inputs and outputs that matter

The output is AUDIO.

• prompt - the sound you want described in words. The default ("Quiet speech and then and airplane flying away") is a good example of the level of detail it likes: name the sound and any sequence of events.
• model - tango2 (default) or tango2-full. tango2-full is the larger checkpoint; start with the default.
• steps (100) - diffusion steps. More steps, cleaner audio, slower and pricier. 100 is already generous.
• guidance (3) - how tightly it sticks to your prompt. Raise it if the output ignores your description; lower it if it sounds forced.

force_rerun forces a new generation when your inputs match a previous run.

How to install it

ComfyUI Manager: search ComfyUI-Replicate, install, restart. Or manually:

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

Restart, and set your Replicate token first:

export REPLICATE_API_TOKEN="r8_************"; python main.py

Grab a token at replicate.com/account/api-tokens. Without it, the node errors on the first run.

Common issues

Here's the honest caveat, and it comes straight from the pack's own README: this pack was built for models that return simple text or image outputs. The README says plainly that when a model "returns audio, video, JSON objects or a combination of outputs, the node will not work as expected," and audio support is listed on the roadmap as a thing still to investigate. Tango returns audio.

The node exists and declares an AUDIO output, so it may well run - but don't be surprised if wiring that output into a save/preview-audio node behaves oddly, or if the result doesn't come through cleanly. If you hit that, it's not you: it's the known rough edge in the pack. Beyond that, the usual suspects apply - a missing REPLICATE_API_TOKEN is the most common hard error, and remember every run costs money on Replicate, so don't leave steps cranked while you're just iterating on a prompt.