Foundation-1 Generate

The model loader gets the attention for auto-downloading 4GB of weights. This node is the payoff: type some tags, pick a BPM, bars, and key, and out comes a sample loop that's actually in time and in key. Foundation-1 (RoyalCities, not to be confused with Stability's own Stable Audio line) was built for exactly this - structured, production-ready loops rather than the "here's a vaguely musical noise bed" you get from generic text-to-audio. If you're a producer who's been fighting audio models to hold tempo, this is the workflow you've been waiting for.

How it works

The node assembles your tags into the full prompt by appending bars, BPM, and key - so Synth Lead, Warm, Wide, Bright, Clean, Melody becomes …, 8 Bars, 140 BPM, E minor. Duration is computed for you from bars and BPM (round(bars × 4 / BPM × 60)), then k-diffusion sampling runs through the stable-audio-tools engine and decodes to a waveform. That's a latent-diffusion audio model doing its thing, same family of math as image diffusion but over mel-spectrogram latents. The pack quietly swaps in its own private k-diffusion for the call, then restores ComfyUI's - you never see it, and it's a genuinely thoughtful shim.

The inputs that matter

• tags - your instrument, timbre, FX, and notation list, comma-separated. The README ships a full tag reference sheet (Master_Tag_Reference.md); trust the tags, not natural language. Don't put BPM, bars, or key in here - those are separate dropdowns.
• bpm (100–150) and bars (4 or 8) - set tempo and loop length; duration follows automatically.
• key - 24 western keys, major and minor. Default is E minor.
• steps (default 250) and cfg_scale (default 7.0) - quality vs. speed. 100–150 steps is the practical sweet spot; guidance 6–8. These behave exactly like their image-model cousins.
• sampler_type - defaults to dpmpp-3m-sde, which the author recommends for quality. dpmpp-2m-sde is slightly faster if you're impatient.
• seed - standard ComfyUI seed control.
• unload_after_generate - offload the model to CPU RAM after each run to free VRAM. On an 8GB card you want this on.
• torch_compile - first run slower, subsequent runs faster. Disable it if you hit errors; it's CUDA-only.
• sigma_min / sigma_max (0.3 / 500) - leave them. These are k-diffusion schedule bounds for people who know what they're doing.

There are two optional inputs worth a look. Connect any AUDIO output (a LoadAudio node, or a previous Generate) into audio and the node switches to variation mode - it re-noises your input and regenerates it guided by your tags, i.e. audio-to-audio remixing. init_noise_level (default 0.7) is the variation strength: 0.1–0.3 stays close to the source, 0.9+ is a creative reinterpretation. This is the feature that makes the pack more than a toy - drop in a rough idea, get back a polished take.

The single output, AUDIO, is a standard ComfyUI audio tensor with its sample rate - wire it into SaveAudio, VHS_VideoCombine, or any audio output node.

Installing and troubleshooting

Same pack as the loader: ComfyUI Manager (search "Foundation-1") or git clone plus python install.py. First run downloads the T5 encoder and weights, so have internet ready. The install traps are the same ones - never pip install stable-audio-tools without --no-deps (pandas 2.0.2 breaks on Python 3.13+), and let the pack manage its own k-diffusion copy.

Realistic gotchas on an 8GB card: out-of-memory mid-generation means turn on unload_after_generate, cut steps to 100–150, and close Discord. Slow generation? pip install sageattention for the attention backend, or use dpmpp-2m-sde. The model caps loops at 20 seconds, which the duration math already respects. And if you're not on NVIDIA - Flash attention makes this pack CUDA-only; forks exist for Mac/CPU, but the official one won't run.

The name overpromises a little - "Foundation-1" sounds like it does one thing, and honestly that's fine. It does one thing really well: in-key, in-tempo loops from a comma-separated sentence.