AceStepTogetherAI_Lyrics

Together AI is a serverless inference platform that hosts a rotating cast of the best open-weight models, and this node is its seating chart. AceStepTogetherAI_Lyrics defaults to meta-llama/Llama-3.3-70B-Instruct-Turbo - a genuinely strong lyricist - and the dropdown also carries Llama 3.1 405B Turbo, DeepSeek-V3, Qwen2.5-72B, Kimi K2, GLM-4-Plus, Mistral Small, and Nous-Hermes-3-70B. The README claims 100+ models are available; like the Fireworks node, the UI only exposes its curated snapshot, but what's listed is a solid spread.

The mechanism

Same family skeleton: text holds your song description, api_key is your Together key (node input, not env var), and the node wraps your text in the shared ACE-Step lyricist prompt before calling Together's OpenAI-compatible endpoint. max_tokens (1024 default), seed, and instrumentality_balance round out the controls; output is a lyrics STRING for your conditioning.

Picking from the list

• Llama-3.3-70B-Instruct-Turbo (default) - well-rounded, fast, and it holds format well. The right first choice.
• Llama-3.1-405B-Instruct-Turbo - the quality pick if you want the big-model polish at serverless convenience.
• deepseek-ai/DeepSeek-V3 - punchy and cheap; good for energetic, pop-structured lyrics.
• Qwen/Qwen2.5-72B-Instruct - reliable generalist, strong in non-English.
• moonshotai/Kimi-K2-Instruct - the wildcard; worth one try just to hear its voice.
• Nous-Hermes-3-70B - the legacy pick, still fine but showing its age.

Where it fits

Together sits in the same tier as Fireworks: paid-per-token serverless access to frontier open models without committing to a subscription. Between the two, Together's default 70B Turbo is a better lyricist out of the box than Fireworks' default - but the two nodes overlap so heavily that most people will pick whichever platform they already have an account on.

Install and gotchas

Pack standard: ComfyUI Manager (search "JK AceStep Nodes"), or

cd ComfyUI/custom_nodes
git clone https://github.com/jeankassio/JK-AceStep-Nodes.git

restart. No extra downloads.

Usual caveats: keys are node inputs (they end up in saved workflow JSON), errors return as the lyrics string, and there's no retry logic for rate limits. Watch token burn on the 405B if you're iterating - the 70B Turbo handles 90% of drafts at a fraction of the cost. It's a solid middle-of-the-family node: better than the free tiers on quality, cheaper than the flagship API nodes on price.