AceStepHuggingFace_Lyrics

Of all the lyrics nodes in this pack, this one has the biggest gap between what its README advertises and what the code actually ships. The README talks up Llama 3.1 405B and Falcon 180B, but the real dropdown - the thing the node sends - contains small 7B-class instruct models: mistralai/Mistral-7B-Instruct-v0.2 is the default, alongside Llama-2-7b-chat, zephyr-7b, falcon-7b-instruct, and similar. Set expectations accordingly: this is a free-ish tier for quick drafts, not your flagship lyricist.

How it works

The node calls HuggingFace's Inference API with your HF token as a Bearer token, wrapped in the same shared ACE-Step lyricist prompt as the rest of the family. Inputs are the usual suspects - text, api_key, model, seed, instrumentality_balance - with one real spec difference: max_tokens defaults to 512 and caps at 2048, the smallest budget in the pack. On a 7B model that's a deliberate guardrail; the small models ramble and repeat if you let them.

What it's good for

Honestly: free experimentation. HuggingFace's inference free tier is limited (and famously rate-limited), but it costs nothing, and for learning the ACE-Step lyrics format or testing whether a song idea has legs, a 7B model's output is fine. Small models are also fast - you'll get a draft in seconds.

What it's not good for: quality lyrics. A 7B model following a long, demanding prompt template will drop the formatting rules, repeat words, and write clichés - the exact things the prompt's vocabulary rules fight against. If a draft sounds generic, that's the model tier, not the node.

Gotchas

• The free tier's rate limits are brutal under any real load. If you're generating a batch, expect 429s.
• Because the free tier only reliably serves small models, the big models in the README (Llama-3.1-405B, Falcon-180B) require a paid Inference provider and aren't even in the dropdown.
• The dropdown's TheBloke/neural-chat-7B-v3-3-GGUF entry is a GGUF file - it's listed in the menu because it exists on the Hub, but it will not behave like a proper API model. Skip it.
• Standard family rules: key is a node input (ships with saved workflows), errors come back as the lyrics string.

Install is the pack standard: ComfyUI Manager (search "JK AceStep Nodes") or git clone https://github.com/jeankassio/JK-AceStep-Nodes.git into ComfyUI/custom_nodes/, then restart.

The blunt take: use this one to prototype for free, then move the winners to the Groq, OpenAI, or Claude nodes for the actual render. It's a nice-to-have member of the family, not the one you reach for when the song matters.