Custom MSS Separate List

What it is

The list-output twin of Custom MSS Separate. Same custom-model scanning, same model_type detection, same audio / model_name / model_type / device inputs and optional params / device_ids / debug sockets - but the outputs are two sockets instead of up to sixteen:

• audios - every stem your model produced, as an AUDIO list.
• stem_names - their names, as a STRING list in the same order.

The names come from training.instruments in the YAML config sitting next to your weights. That's the file that decides how many stems exist and what they're called, so the list is only as correct as your config.

When you'd pick this one

A custom model's stem count is whatever its config says - you might have a four-stem Demucs-family fine-tune today and a two-stem roformer tomorrow. The explicit version of the node reshapes its sockets when the model changes, which means a graph wired to stem_3 breaks the moment you pick something with two stems. The list version doesn't care. Wire audios into a single Save Audio and ComfyUI maps over the list, saving each stem once, with stem_names feeding the filenames positionally.

That's the honest trade: less readable graph, less to rewire. If your workflow only ever uses one stem from one model and you like seeing the name on the socket, use Custom MSS Separate instead.

Setup, in order

1. Clone and install the pack (ComfyUI Manager → Comfy-MSS, or git clone https://github.com/pymss-project/comfy-mss into custom_nodes), then python -m pip install pymss with ComfyUI's own Python.
2. Put each model in its own direct child folder under ComfyUI/models/pymss/custom/, with one weights file (.ckpt, .pth, .pt, .safetensors, .safetensor, .bin) and one .yaml config. .yml won't be picked up.
3. Restart ComfyUI, then click Refresh Models on the node - custom folders are cached in the frontend, so a new folder won't appear until you do.
4. If the run dies with ModelTypeDetectionError, set model_type by hand from the eleven architecture choices instead of auto.

Gotchas

• Multiple files per folder = guesswork. Search order is: same-stem name pair, else first weights file plus first YAML alphabetically. One model per folder, always.
• VR/UVR models are deliberately excluded. This node is for MSST architectures; legacy VR .pth files belong to VR Separate, and giving one a YAML here won't sneak it in.
• Lists are stems, not songs. One input audio, several outputs.
• No catalogue, no downloads. There's no download_missing or source here for a reason; the weights are yours.
• v1.0.x workflows need rebuilding - node types were renamed in v1.1.0.