AudioX Text to Music

Let's be straight about what this node is: it's AudioX Text to Audio with a music-shaped default prompt. Same model, same pipeline, same generate_diffusion_cond call, same 250-step dpmpp-3m-sde sampling. The difference is the default text (A music with piano and violin), a prompt-enhancement pass that nudges the model toward musical context, and the framing in your head.

That's not a knock. If you want a quick ambient bed or a demo loop without hunting through the enhanced node's style dropdowns, this is the lean one to reach for. Just understand what you're getting: the AudioX finetune is a text-conditioned audio model, not a purpose-built music generator like a dedicated music diffusion checkpoint. It'll happily produce something piano-and-violin-ish for a 10-second loop. It will not reliably give you song structure, a verse-chorus-verse arc, or a 3-minute composition - that's beyond what this weight does, and no prompt tweaking changes that.

The inputs

• model - from the AudioX Model Loader (AUDIOX_MODEL).
• text_prompt - describe the music. Being specific helps: instruments, tempo words, mood.
• steps - 250 default. Drop to 50–100 for iteration.
• cfg_scale - 7 default (0.1–20). Higher sticks to the prompt harder.
• seed - -1 for random, or lock it to reproduce a take.
• duration_seconds - 1–30, default 10.

Output is audio (AUDIO), the same {waveform, sample_rate} dict every AudioX generator returns. Pipe it into the pack's Advanced Volume Control for loudness-matching, or a save node to write the file.

Install and gotchas

Same as every node in this pack - the full ritual lives in the AudioX Model Loader article, but the short version:

cd ComfyUI/custom_nodes
git clone https://github.com/lum3on/ComfyUI-StableAudioX.git
cd ComfyUI-StableAudioX
pip install -r requirements.txt

Plus the AudioX weights and config.json from HKUSTAudio/AudioX on Hugging Face in ComfyUI/models/diffusion_models/ (weights renamed to AudioX.ckpt).

Two things bite people here specifically:

• "Music" results that sound like foley. The enhancement logic adds the word "musical" to the prompt if it isn't there, but the model only has so much genre knowledge. If you want actual control, skip this node and use AudioX Enhanced Text to Music, which layers style/tempo/mood terms onto the prompt properly.
• VRAM. Defaults are heavy. This pack's own README admits it was tested on a single 4090 with 16GB+ VRAM in practice. If you're on less, start with duration_seconds: 5 and fewer steps.

It's a two-node workflow - loader into this, then out to a saver - which makes it the pack's best "does this even work" test. Run it once, be amused, then reach for the enhanced version when you want actual control.