AQ_Gemini_acstep15

This is the node at the center of this pack's whole reason for existing. The author's own description of AQnodes is "mostly for llm calling and automatizing pulid multiface" - and the acestep family is the LLM half. Point this node at a song idea and it returns structured metadata ready to feed ACE-Step (the open-source diffusion music model - think of it as a text-to-audio Stable Diffusion for full songs), plus cover-art direction and YouTube upload text. If you've seen those "generate a full song, cover, and upload it" workflows, this is the brains.

What it produces

AQ_Gemini_acstep15 is AQ_Gemini, but opinionated. Instead of freeform text it forces the model to return a fixed JSON schema:

• tags - a music style prompt (the primary control for genre, mood, instrumentation)
• lyrics - with ACE-Step's structure tags ([verse], [chorus], [inst]) and language codes
• seed, bpm, durationSeconds, timesignature, language, keyscale
• coverImage - a 150–250 word art-direction prompt for the album cover
• title, tagline, headline - text for the cover / thumbnail
• yt_description, yt_tags, yt_title - a full YouTube packaging kit

The craftsmanship is in the schema descriptions: the fields literally tell the model how to write an ACE-Step style prompt, how to format lyrics for its text encoder, typical BPM ranges by genre, and what a keyword-rich YouTube title looks like. That's why the output is so usable - the model is being instructed on the pipeline it's feeding.

How it works

The node streams generate_content via the google-genai stack. For non-gemma models it sets response_mime_type: application/json plus response_schema, so the API returns valid JSON that matches the schema. For gemma models it falls back to a prompt-level instruction. Then it parses the response and coerces every field to a typed output with sensible defaults: bpm→120, duration→120.0s, keyscale→C major, timesignature→"4", language→"en". The outputs are typed so they plug straight into downstream nodes - seed and bpm are INTs, duration a FLOAT, timesignature/language/keyscale COMBOs, everything else STRING. No API key → it returns empty strings with those defaults instead of erroring.

Inputs

gemini_api_key, model_selection (gemma-3 family, gemini-flash-latest/lite, 2.5 flash/pro, gemini-3 previews, or custom), custom_model, prompt (your song idea), system_message, temperature, top_p, seed (pass a seed to vary or reproduce metadata), and an optional image (a reference art to inspire the cover).

Outputs

The full 17: 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

Gemma models get prompt-based JSON rather than schema-enforced, so it's less reliable there - for strict output, pick a gemini model. And this is one node in a family because LLMs differ: AQ_OpenAI_acstep15 uses OpenAI's Responses API, AQ_openai_compatible works against any OpenAI-compatible endpoint, and AQ_parse_json_to_acestep turns arbitrary JSON into the same slots. If you're using any of them, the outputs feed ACE-Step's audio text encoder for the music, AQ_CoverPrompt for the cover, and an upload node for YouTube.