AceStepGroq_Lyrics
Groq's Free Tier for Rapid Song Drafting
- lyrics
Groq runs open-weight models on custom inference hardware that's absurdly fast, and - importantly for anyone not paying per call - its free tier is real. AceStepGroq_Lyrics is the node in this pack that turns lyric generation into an interactive workflow: type a style, hit run, and you get a full ACE-Step-formatted lyric sheet back in a couple of seconds. When you're iterating on song ideas that's a genuine advantage over the ChatGPT/Claude nodes, which feel ponderous by comparison.
The mechanism
Same shared lyricist framework as the rest of the family - the node wraps your description in the production-tag prompt (short lines, tags on their own lines, vocabulary rules) and returns lyrics ACE-Step can actually interpret. It calls Groq's OpenAI-compatible endpoint with a style input (not text - "Music style or style prompt" is the placeholder), your api_key, model, max_tokens, seed, and instrumentality_balance. Output is the standard lyrics STRING for your conditioning.
The one real spec difference: max_tokens runs up to 8192 here, versus 4096 on most siblings. Probably not a knob you'll need for a song, but it exists if you ever want the model to write a whole album's worth in one call.
Picking a model
The dropdown defaults to llama-3.3-70b-versatile, which is the quality pick and a good default for lyrics. llama-3.1-8b-instant is the speed demon - genuinely usable for rapid A/B drafts, and it's the model I'd use for throwing twenty style variations at the wall. The other entries are worth a glance: openai/gpt-oss-120b is a strong newer option that punches well above its cost, and the llama-4 scout/maverick entries are the newest but still settling. llama-guard-3-8b is a safety classifier model - you don't want it generating lyrics, it's in the list because it's on Groq's menu.
Where it shines and where it doesn't
Shines: free, instant, great for high-volume drafting and for testing whether an idea is worth a paid render. Doesn't: open-weight models at this speed tier are not the last word in poetic lyricism - if you want the best-written output and cost doesn't matter, the Claude node is the one. Think of Groq as the fast sketchpad in front of a slow, careful render.
Install and gotchas
Same pack install: ComfyUI Manager (search "JK AceStep Nodes") or
cd ComfyUI/custom_nodes
git clone https://github.com/jeankassio/JK-AceStep-Nodes.git
then restart. Key caveats: the free tier rate-limits hard under bursts (fine for hobby use, painful if you queue a big batch), and - as with every node here - your pasted key lives in the node and travels with saved workflows, so don't share workflow JSON that contains it. Errors come back as the lyrics string; read the node output before debugging the sampler.
Inputs (6)
| Name | Type | Default | Description |
|---|---|---|---|
| style | STRING | — | |
| api_key | STRING | — | |
| model | COMBO | llama-3.3-70b-versatile | 9 options: llama-3.3-70b-versatile, llama-3.1-8b-instant, llama-3.2-1b-preview, llama-3.2-3b-preview, openai/gpt-oss-120b, openai/gpt-oss-20b, +3 |
| max_tokens | INT | 1024256–8192 | — |
| seed | INT | 00–18446744073709550000 | — |
| instrumentality_balance | COMBO | Balanced | 5 options: Pure Instrumental, Very Instrumental, Balanced, Vocal Focused, Vocals Only |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| lyrics | STRING | — |