TangoFluxSampler

TangoFluxSampler is the node that does the actual magic: you type "a dog barking near the ocean, ocean waves crashing" into a box and it returns latent audio, at up to 44.1kHz and up to 30 seconds, in roughly the time it takes to pour a coffee. The model's whole pitch is speed - the arXiv paper is literally titled Super Fast and Faithful - and this is the node that delivers it. It's the middle of a three-node chain: TangoFluxLoader → this → TangoFluxVAEDecodeAndPlay. Your job is the prompt and a handful of sliders.

Text-to-audio is a thinner corner of the ecosystem than image or video generation - most people's first experience of it is Veo-style video that happens to make sound. TangoFlux is the open, standalone alternative, and it's a genuinely good ambience-and-sound-effects machine. What it isn't: a precise Foley generator. A reviewer on the pack's launch thread put it bluntly - a hammer hitting a table doesn't sound like the same table on every hit, and melodic/rhythmic noise turns into "random midi triggers strung together." For atmospheric layers, crowd noise, wind, rain, synth-ish textures, it's great. For sound design where each hit has to match, it'll disappoint. The author's own fix for the worst of it: run 50 steps, not 25.

How it works

The prompt goes through the FLAN-T5-large text encoder. With guidance_scale above 1, it does classifier-free guidance - encoding an empty prompt too and interpolating between the two, batched into one pass to stay fast. Then a flow-matching Euler scheduler walks the latent timeline over steps, and you get back a TANGOFLUX_LATENTS blob (the latent audio is ~645 frames, which the VAE later expands to 44.1kHz waveforms). It's the same denoising intuition as image sampling, just on sound.

The inputs you'll actually set

• prompt - multiline, and dynamic prompts are on, so [thunder|rain] storm wildcards work.
• duration - 1 to 30 seconds, default 10. Longer is slower and eats VRAM.
• steps - default 50. This is the quality knob; the author's own recommendation when output sounds rough is exactly this.
• guidance_scale - default 3. Lower for more literal, higher for more "faithful to prompt" - in practice 2–5 is the zone.
• seed, batch_size - standard. One honest warning: the model code carries the comment "Only tested for single inference," so treat batch_size above 1 as untested territory.
• offload_model_to_cpu - for low-VRAM cards; it shunts the model to CPU after sampling to free room for the VAE decode. Slower, but turns OOMs into working runs.

Output

a single latents output, which feeds TangoFluxVAEDecodeAndPlay. That's the whole graph - nothing else accepts TANGOFLUX_LATENTS, so you can't misroute it.

Installing

is the pack-wide step, so if you've already got the pack you're done:

cd ComfyUI/custom_nodes
git clone https://github.com/LucipherDev/ComfyUI-TangoFlux
cd ComfyUI-TangoFlux
python install.py

(Or ComfyUI Manager → search "TangoFlux".) That installer pulls the TangoFlux weights plus the T5 encoder - a few GB, one time.

Troubleshooting

Output sounds muddy or falls apart → crank steps to 50 and make sure it was there to begin with; that's the documented fix. VRAM errors at decode time → flip offload_model_to_cpu on. Everything works but it's slow → the speed lever is TeaCache, and it's on the loader node, not here - enable_teacache with rel_l1_thresh around 0.25 roughly halves sampling time at modest quality cost. And if you're chasing precise one-shot sound effects, lower your expectations before you lower your settings - this model's ceiling on that is a feature of the architecture, not a slider you missed.