MusicMaker ONE STUDIO (TJ)

MusicMaker ONE STUDIO (TJ) is the audio corner of designloves2's "ONE STUDIO" family, and it's the most Suno-shaped thing in the pack. Where the sibling nodes wrap image or video models, this one wraps two local music engines - Ace-Step 1.5 (the default) and MiniMax Music 3 - behind a single node that looks like the SUNO web app: lyrics on the left, a playlist on the right, an inline player along the bottom. No API calls, no key for the music itself. It's flagged 🧪 experimental, so UI details may move - the engines are real.

What it actually is

Every node in this pack shares a philosophy: instead of wiring a graph, you work in a studio UI that is the node, which silently assembles the real ComfyUI graph for you. That pays off hardest here - a music graph is a mess to hand-build, and MusicMaker hides all of it.

The engines differ under the hood. Ace-Step 1.5 runs a DualCLIPLoader(ace) + AuraFlow setup with a 3-stage progressive SamplerCustom refine (30 steps at CFG 0, then 20, then 15) on the JKASS quality sampler, and takes BPM, key, time signature and language as structured inputs at 48 kHz. MiniMax Music 3 is native to recent ComfyUI core (MiniMaxMusic3TextEncode and friends, no extra packs) and folds BPM/key/vocals into the caption text instead.

How generation actually works (the part that surprises people)

You do not hit ComfyUI's RUN to generate. It's true of every node in this pack and the first thing that trips people up: there's a big ▶ Generate button inside the node, and that's the trigger. Clicking it makes browser-side JS snapshot the whole compose panel and submit a complete, self-contained graph to the server queue - one graph per job, drained FIFO by a single worker, with per-card cancel and a Stop-all. Because each job is a full graph submitted at once, it keeps running server-side even if you close the tab - unlike the pack's MiniMax H3 video node, which needs the browser open to feed the next clip.

That's why MusicMakerOneTJNode is such a thin shell. It takes no required inputs, exposes one optional pipe socket (a TJ_PROMPT_PIPE from the author's ComfyUI-TJ_NODE PromptDB loader - present fields override the node's settings at generation), and its single audio output just hands back the most recently generated track so it can flow onward in your graph. The output slot is for interoperability; day to day, the playlist inside the node is the product.

The controls that matter

The compose panel is where you spend your time:

Engine - Ace-Step 1.5 or MiniMax Music 3, top of the node.
Song / Instrumental toggle - flipping to instrumental hides the lyrics and vocals sections entirely.
Lyrics with [Verse]/[Chorus] structure - or leave it bare and let the ✨ button write it.
Style caption, plus BPM / key / time signature / length.
The ✨ LLM does a huge share of the work - style captions, full lyrics, even the Krea2 album-cover prompt. Backend is chosen in Settings: local GGUF via ComfyUI-TJ_NODE, cloud OpenRouter, or ComfyUI's own TextGenerate.

The lyric writer is genuinely decent about length: it gets the real target duration, so a three-minute song doesn't come back as a twenty-second verse. Finished tracks land in the playlist with inline play, per-track Info, Reuse, and a tagged-MP3 download (title, artist, cover, lyrics via ffmpeg).

Installing it

Install the pack once - via ComfyUI Manager (search ComfyUI-TJ_NODE_STUDIO_ONE) or:

cd ComfyUI/custom_nodes
git clone https://github.com/designloves2/ComfyUI-TJ_NODE_STUDIO_ONE.git

Then the step the README can't stop yelling about: Manager installs this pack's Python requirements but not the sibling node packs it depends on. Right after install, run the included install_requirements.bat (Windows) or bash install_requirements.sh (Mac/Linux) once and restart ComfyUI. MusicMaker specifically needs JK-AceStep-Nodes (its Ace-Step default sampler is jkass_quality), plus optionally ComfyUI-TJ_NODE for the local-GGUF LLM path and ComfyUI-Openrouter_node for cloud. A startup banner tells you what's missing.

Models you need

Nothing auto-downloads. Ace-Step 1.5 and MiniMax Music 3 each ship their own diffusion model, text encoder(s) and audio VAE - grab whichever engine's files you want, drop them in models/ (diffusion_models / text_encoders / vae), then pick them in the node's ⚙ Settings. Recent ComfyUI is required for the MiniMax engine.

Where people get burned

Missing JK-AceStep-Nodes is the classic failure: the Ace-Step engine loads but generation dies because jkass_quality doesn't exist. Install it, or switch to another sampler in Settings. Vocals on the MiniMax engine can sound "windy" - a model trait (its audio VAE renders 5–12 kHz as broadband noise), not your cfg. And set expectations honestly: the community consensus on Ace-Step - strong instrumentals, weak-to-middling generated vocals and lyrics - applies here too. This is the best local ComfyUI path to a finished track; it is not Suno.