FL Song Gen Model Loader

Every FL SongGen workflow starts here. This node loads a SongGeneration variant and hands a SONGGEN_MODEL to the Generate, Style Transfer, and Auto Style nodes. There's no trick to it - but the choice you make in it decides how long your songs can be, how much VRAM you need, and whether your English lyrics actually work. The defaults are sensible, so this is a "set once and forget" node. The part that surprises people is the download.

The variant choice is the whole node

model_variant is a dropdown, and it matters more than anything else in the pack:

| Variant | Languages | Max | VRAM (normal / low) | |---|---|---|---| | songgeneration_base | Chinese only | 2m30s | 16 / 10 GB | | songgeneration_base_new | Chinese + English | 2m30s | 16 / 10 GB | | songgeneration_base_full | Chinese + English | 4m30s | 18 / 12 GB | | songgeneration_large | Chinese + English | 4m30s | 28 / 22 GB | | songgeneration_v1_5_beta | + Spanish, Japanese | 4m30s | 24 / 16 GB |

The default is songgeneration_base_new, and the author's own guidance is to keep it for English lyrics - it's the quality/speed/VRAM sweet spot. Reach for base_full or large when you actually need the 4m30s cap, and treat v1_5_beta as what it says: experimental.

The rest

• memory_mode - auto checks your free VRAM and picks for you; normal is fast but greedy, low trades speed for VRAM, ultra gets you down to roughly 6 GB and is the slowest. auto is genuinely the right default.
• force_reload - default false, and leave it there. Flip it to true only if you switched variants and it's serving you a stale cached load from the in-session cache.

The single output, model, feeds any of the three generation nodes.

How the download actually works

On first use the node auto-downloads the weights to ComfyUI/models/songgen/ - from a runtime repo (lglg666/SongGeneration-Runtime) for the checkpoints plus the variant-specific repo, and it also pulls the Demucs stem-separation model used by style transfer. So the first load is a multi-GB download that just appears in your terminal, and the console logs tell you exactly which model it's fetching. The source also bundles the code the model needs (in codeclm/third_party), which is why you only download weights.

Install

ComfyUI Manager (search "FL Song Gen"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI_FL-SongGen.git
cd ComfyUI_FL-SongGen
pip install -r requirements.txt

Restart ComfyUI after. Requirements are heavy but largely shared with modern ComfyUI: torch, torchaudio, transformers, diffusers, librosa, openunmix, descript-audio-codec, k-diffusion, x-transformers. The pack bundles a patch for AMD GPUs so the audio-codec import doesn't crash on torch.distributed.ReduceOp.

Troubleshooting

• It's downloading forever on first run - normal. Several gigabytes across checkpoints and the Demucs model.
• Model files not found - if the auto-download ever fails, the node's error message points you at the HuggingFace repos; expected location is ComfyUI/models/songgen/<variant>/ with config.yaml and model.pt.
• OOM when generating - drop memory_mode to low or ultra before blaming the generation node. 16 GB RAM is the floor; 32 GB is the "stop thinking about it" level.
• "FL Song Gen" not showing in Manager - update ComfyUI Manager's database and restart; then it's findable by the pack title.