AQ_openai_compatible

The flexible one in the acestep family. Same fixed song-metadata schema - tags, lyrics, bpm, duration, keyscale, cover art, YouTube text - but instead of OpenAI's Responses API it posts to whatever endpoint you give it, defaulting to OpenAI's chat-completions URL. Point it at Ollama, vLLM, LM Studio, Together, Groq, or any local OpenAI-compatible server, and you get the same structured output without the cloud dependency.

Why you'd pick this over the others

If you want the whole song-metadata pipeline but refuse to send your lyrics prompts to a cloud API - or you just already run a local LLM - this is the bridge. It uses chat completions with a strict JSON-schema response_format, which most OpenAI-compatible servers support these days. The trade-off is real: strict schema enforcement (additionalProperties: false) is only as good as the server, and weaker local models won't always honor it, so expect the occasional retry.

How it works

Same coerce-and-split as the siblings: it sends temperature/top_p/seed (omitted when negative), parses the JSON, and splits it into typed outputs with defaults (bpm→120, duration→120s, keyscale→C major). An optional image becomes a data:image/png;base64 URL in the message. The schema descriptions carry the full ACE-Step flavor - tags as a music style prompt, lyrics with [verse]/[chorus]/[inst] structure tags and language codes, BPM ranges by genre - so the model is being instructed on the exact pipeline it feeds.

Inputs

openai_api_key, model (9 choices, gpt-5-nano default), custom_model, prompt, system_message, temperature, top_p, seed, verbosity, reasoning_effort, api_endpoint (default https://api.openai.com/v1/chat/completions), and an optional image.

Outputs

The 17-slot acestep set: response_json, tags, lyrics, seed, bpm, duration, timesignature, language, keyscale, coverImage, title, tagline, headline, yt_description, yt_tags, yt_title, instruction.

Install

Part of AQnodes:

cd ComfyUI/custom_nodes
git clone https://github.com/2frames/ComfyUI-AQnodes
cd ComfyUI-AQnodes
pip install -r requirements.txt

or search "AQnodes" in ComfyUI Manager and restart.

Gotchas

Not every local server honors strict JSON schema - Ollama and vLLM do; older ones return markdown-wrapped JSON and the node raises when parsing fails, so keep an eye on the console. The verbosity/reasoning_effort controls inherited from the OpenAI twin map loosely to the chat-completions payload and may be ignored by local models. And note the default endpoint is OpenAI's chat completions URL - it works if you want OpenAI after all, just with a different API shape than AQ_OpenAI_acstep15's Responses API.