SCG Foundation-1 Loader

Foundation-1 (by RoyalCities) is a text-to-sample music model - you describe a loop like "warm synth lead, bright, melody" and it generates tempo-synced, key-aware audio. It's built on Stability AI's stable-audio-tools library and inherits its Community License quirk: free for non-commercial work, or commercial use under $1M revenue. The SCG Foundation-1 Loader is the mandatory first node in any workflow from this pack, because it does the two jobs nobody wants to do by hand: pulling the ~1.5 GB weights from HuggingFace and getting them onto your GPU.

Everything else in the pack - Prompt Builder, Random Prompt, Sample Generator - assumes you've run this node once. Wire its model output straight into the Sample Generator's model input and you're done. It's a one-wire setup.

What it does

On first run the Loader downloads Foundation_1.safetensors plus the model config from the RoyalCities/Foundation-1 HuggingFace repo and caches them at ComfyUI/models/audio_checkpoints/Foundation-1/. It also pulls the T5 tokenizer on that first run. Later runs just load from disk - no manual model-file hunting, no symlink dance into a checkpoints folder.

Under the hood it's unglamorous: build the model from config, load the checkpoint, move it to whatever device you picked, and hold it in memory. The real work lives in stable-audio-tools, which the pack installs for you.

The two inputs that actually matter

• device - auto (default), cuda, or cpu. Auto picks CUDA if it's available and falls back to CPU. CPU technically works (the pack even ships a flex_attention shim so older PyTorch doesn't crash on import), but you'll be waiting a long while on a diffusion loop. This is a GPU node in practice.
• keep_model_loaded - true by default. Leave it on: the model stays warm in VRAM so every subsequent generation is fast. Flip it to false if you're juggling VRAM with image models and want the model unloaded after each run.

The single output, model (typed F1_MODEL), feeds the Sample Generator. Nothing else consumes it, and there's no other wire to pull.

Installing the pack

Use ComfyUI Manager and search for the pack title scg_Foundation-1-comfyUI (or just "SCG Foundation-1"). Manager runs install.py for you, which is the install method you should prefer. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/SanDiegoDude/scg_Foundation-1-comfyUI
cd scg_Foundation-1-comfyUI
python install.py

Here's the one gotcha that bites everyone, and it's worth saying loudly: do not run pip install -r requirements.txt by itself, and never pip install stable-audio-tools normally. stable-audio-tools pins dozens of training-only dependencies to ancient versions (torch, numpy 1.23, pandas 2.0…) and will wreck your ComfyUI environment. install.py handles the --no-deps install plus the handful of lightweight inference deps. If you must pip manually, it's two steps in order: pip install stable-audio-tools --no-deps, then pip install -r requirements.txt.

Troubleshooting

If you see a loud banner at startup saying "stable-audio-tools is NOT installed," the nodes will still appear in the menu but won't generate. Fix: run python install.py from the node folder, or pip install stable-audio-tools --no-deps. And budget VRAM - generation wants roughly 7 GB, so an 8 GB card is the practical floor. If ComfyUI is running out of memory, keep_model_loaded = false plus device = auto is the first thing to try.