AudioX Enhanced Text to Music

AudioX Text to Music asks you to put everything in one sentence. This node hands you dropdowns instead: pick a music_style, a tempo, a mood, and it composes a much better prompt than you were going to type. It's the pack's answer to "I want something jazzy and mysterious at a slow pace" without having to know how to say that to a diffusion model.

Same backend, same sampler, same model - the "controls" are prompt construction. music_style: jazz prepends "jazz music, improvised"; tempo: slow appends "slow tempo, relaxed pace"; mood: mysterious appends "enigmatic, suspenseful, dark". The node also makes sure the word "music" is in the prompt (the model genuinely performs better when you tell it it's music). Then the same enhancement machinery as the enhanced text-to-audio node takes over: conditioning modes, adaptive CFG, and the four-way standard/enhanced/super_enhanced/multi_aspect ladder.

Inputs that matter

• model - from the AudioX Model Loader (AUDIOX_MODEL).
• text_prompt - your seed idea ("a peaceful piano melody" by default).
• music_style (optional) - classical, jazz, electronic, ambient, rock, folk, cinematic, none.
• tempo (optional) - slow, moderate, fast, very_fast, none.
• mood (optional) - happy, sad, peaceful, energetic, mysterious, dramatic, none.
• negative_prompt (optional) - same honesty note as its sibling: the tooltip says "currently logged only - implementation pending." It doesn't veto anything yet.
• enhance_prompt / conditioning_mode / adaptive_cfg - same as the enhanced audio node.

Output is audio (AUDIO), and it flows into the same volume-control / save pipeline as everything else in the pack.

Install & gotchas

Pack ritual applies:

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

AudioX weights + config.json from HKUSTAudio/AudioX on Hugging Face in ComfyUI/models/diffusion_models/ (weights renamed AudioX.ckpt), plus system ffmpeg/libsndfile.

The realistic picture:

• These are prompt hints, not synthesis parameters. "Rock" doesn't switch a model mode - it changes a few words in the text conditioning. Expect hints of the style, not a faithful genre performance. That's inherent to how this model works, not a bug in the node.
• The negative prompt does nothing yet. Same pending implementation as every enhanced node in the pack. Don't build a workflow around it.
• Tempo is fuzzy. The model doesn't quantize to a BPM; "fast" is vibes. If you need exact timing, generate and then time-stretch in a real audio tool.
• VRAM and first-run lag remain the same walls as the rest of the pack - ~16GB VRAM territory per the README, and a slow first import.

For quickly sketching "what does this video feel like as electronic music," this node is genuinely fun - three dropdowns beat a paragraph of prompt engineering. Just remember you're steering text, not operating a synth.