AudioX Enhanced Text to Audio

The basic AudioX Text to Audio node takes your sentence and runs with it. This one gives you the steering wheel: prompt templates, style modifiers, an "adaptive CFG" that tweaks guidance based on how specific your prompt is, and four conditioning modes. Same AudioX diffusion backend under the hood - this is a prompt-and-conditioning control surface on top of the exact same dpmpp-3m-sde sampler.

Here's the honest mechanic: most of the "enhancement" is careful string surgery before the model ever sees your text. Pick a prompt_template and it appends a preset phrase ("ocean waves, water sounds, peaceful seaside ambience"). Pick a style_modifier like cinematic and it prepends "cinematic, film-quality". Turn on enhance_prompt and it runs the same auto-enhancement as the basic node, adding audio-domain keywords. super_enhanced mode stacks all of that plus keyword expansion and term emphasis; multi_aspect builds multiple conditioning vectors. None of it changes the model - it changes what gets typed into the text conditioner, which is exactly how you steer a text-conditioned audio model.

Inputs that matter

• model - from the AudioX Model Loader.
• text_prompt - your core description.
• steps / cfg_scale / seed / duration_seconds - the standard four (250 / 7 / -1 / 10).
• negative_prompt (optional) - the one caveat: the tooltip on this field says "currently logged only - implementation pending," and the code confirms it just logs the value. Negative prompting in this pack is not yet real; fill it in for future-proofing, not for results.
• prompt_template (optional) - music_ambient, nature_forest, urban_traffic, action_impact, and five more, plus none.
• style_modifier (optional) - cinematic, realistic, ambient, dramatic, peaceful, energetic, none.
• conditioning_mode (optional, default enhanced) - standard (no enhancement), enhanced, super_enhanced, multi_aspect.
• adaptive_cfg (optional, default true) - measures prompt specificity and nudges CFG accordingly.

Output is audio (AUDIO), wired the same as every generator in the pack.

Install & troubleshooting

Standard pack install - clone, install, download weights:

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

Weights + config.json from HKUSTAudio/AudioX on Hugging Face into ComfyUI/models/diffusion_models/ (weights renamed AudioX.ckpt); ffmpeg/libsndfile at system level.

Where people get tripped up:

• "My negative prompt does nothing." Correct - see above. It's logged but unimplemented. This is the pack's biggest misleading label, and it's worth knowing before you burn an hour testing it.
• Over-enhancement. super_enhanced + multi_aspect + templates can produce a prompt so padded that the model flattens out. If results get generic, back off to enhanced mode and a clean prompt.
• First run lag + VRAM. Lazy heavy imports on first queue; README's honest spec is ~16GB VRAM on a 4090. Keep duration_seconds modest on weaker cards.
• Adaptive CFG surprises. If results jump around when you edit your prompt, it's adaptive_cfg rescaling guidance on you. Turn it off for A/B comparisons.

This is the version of text-to-audio you'll actually want day to day once you know the pitfalls: the templates are handy for ambience beds, and the style modifiers do real work on foley. Just remember the negative prompt is a promise, not a feature - yet.