Creativity

Music generation is still the weird edge of ComfyUI - node packs rather than anything near your checkpoints, instrumentals fine, vocals the thing everybody argues about. Plenio is an ambitious attempt to fix the packaging problem: a whole song pipeline - brief, lyrics, score, render, master, export - built on ComfyUI's native YuE2 and MiniMax Music 3 nodes.

Which brings us to PlenioCreativity, display name Creativity. It is not a model, not a sampler, and it won't write you a better song. It translates the one number you actually want to think about - the creativity slider on the Song Brief - into the four sampling fields that control what happens. If you've seen a node called "Take creativity" with three numeric outputs and wondered what it does, this is that answer.

What it's for

YuE2 plans a song in two passes: one writes the score as ABC notation, the other performs that score into music tokens. Both are samplers taking temperature, top_p and top_k; the writer model drafting your lyrics takes a temperature too. Four stages, one editorial intent - so Plenio exposes one slider and lets a node do the arithmetic.

You set creativity once on the brief and this converts it per stage. Doing it by hand means remembering that a plan at 0.55 sounds one way and a take at 1.05 another - a table nobody wants in their head while writing a chorus.

How it works

Three anchors per stage - 0, 50, 100 - with a straight line between them. 50 is deliberately the boring value: it reproduces exactly what the templates hardcoded before the slider existed. For the plan that's 0.70 / 0.90 / 30; for the take, 1.00 / 0.95 / 100. Those are YuE2's own upstream defaults, which is the point - 50 means untouched.

The ends stay nowhere near the extremes. At 100 the plan reaches 0.92 / 0.95 / 50 and the take 1.10 / 0.97 / 150, under the range upstream warns about. Critically, the repetition penalties never move: music lives on repetition, and a penalty below 1 turns a song into an infinite loop. The writer and arrangement stages only interpolate a temperature (0.60→0.80→1.00 and 0.55→0.70→0.90); their top-p and top-k outputs exist but are inert there.

It also prints its resolved numbers onto itself every run: "Creativity 75 (varied) - take: temperature 1.05, top-p 0.96, top-k 125".

Inputs and outputs that matter

Three inputs:

• stage - writer, arrangement, plan, take. The one you'd set by hand, and only if you're building a block yourself; the shipped templates already have a copy per block (Writer creativity, Plan creativity, Take creativity).
• creativity - the 0–100 slider, default 50.
• brief (optional) - connect the Song Brief here and its creativity wins. That's the trap: the widget on the node is ignored once a brief is plugged in, so don't sit there dragging it wondering why nothing changes.

The outputs are temperature, top_p and top_k. temperature goes wherever the stage needs it: ComfyUI's text-generation node or Plenio's local LLM node for the writer, YuE2GenerateABC for the plan, YuE2GenerateMusic for the take. top_p and top_k only reach the two YuE2 nodes - text generation doesn't expose them the same way.

Installing it

The node ships with the pack, so nothing separate to fetch:

cd ComfyUI/custom_nodes
git clone https://github.com/jplenio/Plenio-Music-Production-System

Or in ComfyUI Manager, search Plenio Music Production System; CLI users get comfy node install comfyui-plenio-music.

You need ComfyUI 0.37.0 or newer - the pack is a thin layer over native YuE2, SheetSage2, MiniMax Music 3 and text-generation nodes, and older builds give you a graph full of red. requirements.txt genuinely installs nothing; the only optional dependency is faster-whisper, for cover lyrics. The weights are the heavy part: YuE2 is CC BY-NC and the int8 checkpoint alone is about 4 GB, so 8–12 GB of VRAM means short songs. ComfyUI's missing-model dialog gets them.

Where people get tripped up

You can't find the node. It's inside subgraphs - the Plenio blocks are collapsed groups. Right-click a block and Expand, or open it with the icon at its top right.

Your old workflow ignores the slider. Workflows saved before 1.0.2 hardcode their sampling in the YuE2 blocks. The slider still changes the writing, but the render won't follow until you start from a current template and copy your brief across.

Higher isn't better, it's riskier. Above 60, expect more misses along with the surprises - new take seed, listen to a few. And the plan stage goes inert once a finished score reaches YuE2 (an edited score, a creative mode, the DAW template): the score already is the plan, so only the take's sampling still counts.