SongBloom Model Loader

Every SongBloom workflow starts here, because this is the node that turns an empty ComfyUI into a machine that can write a song. SongBloom is Tencent AI Lab's open 2B-parameter music model - the closest thing the local scene has to a Suno, and this loader is the front door to it. Nothing else in the pack runs until this node finishes, and "finishes" can mean a while: the weights are around 8GB, and if you've never loaded them, the first run downloads them to disk.

What it actually does is decide between two jobs: grab the weights from ComfyUI/models/SongBloom/ if they're already there, or pull them from HuggingFace (CypressYang/SongBloom) when they aren't. Then it builds the model from the config file, wires up the VAE, and hands you a model object the rest of the pack consumes.

The inputs that matter

Only four, and two of them do most of the work:

• model_name - songbloom_full_150s or songbloom_full_150s_dpo. Pick the DPO variant. The community consensus is that the DPO fine-tune sounds noticeably better, even if it clearly got overtrained on Chinese pop with a female lead.
• dtype - float32 (default) or bfloat16. This is your VRAM lever. Under about 8GB, bfloat16 is the difference between "works" and "CUDA out of memory."
• load_mode - local_first (default), local_only, or download_only. local_only is handy if you've already got the files and don't want the node touching the network; download_only forces a fresh pull.
• lyric_processor - pinyin, phoneme (default), or none. This drives how your lyrics get turned into something the model can sing. phoneme is the general choice; pinyin helps with Chinese; none skips conversion entirely.

force_reload is the only optional input - set it true to force a re-download of the model files when a download went sideways.

Output is a single model (SONGBLOOM_MODEL), which feeds straight into SongBloomGenerator or SongBloomBatchProcessor.

Installing the pack

ComfyUI Manager users: search "SongBloom_ComfyUI" and install. Manual path:

cd ComfyUI/custom_nodes
git clone https://github.com/xuchenxu168/SongBloom_ComfyUI
cd SongBloom_ComfyUI
pip install -r requirements.txt

Restart ComfyUI after that. The requirements list is heavy - torchaudio, lightning, transformers, omegaconf, descript-audio-codec, spacy, nltk, and the g2p stack. If you skip the pip install step, the loader just fails with "SongBloom dependencies not available," which is the pack's most common first-run error. There's also an optional python install.py that checks your environment and sets up the model directories.

Where people get burned

• The first load is slow and eats bandwidth. The auto-download is ~8GB. If it fails, check your HuggingFace access and retry with force_reload on.
• VRAM. SongBloom wants 8GB for float32, 6GB for bfloat16 per the README - and the community would tell you those are optimistic. If you get "CUDA out of memory," the fix order is: bfloat16 → fewer steps downstream → reduce max_frames.
• The OmegaConf "resolver 'eval' is already registered" error when you reload or switch models. That's a known multi-load bug, patched in v1.1.0; if you still hit it, restart ComfyUI.
• Mind the license. SongBloom's weights are academic/non-commercial. Fine for experiments, not for release-day commercial distribution.

One last honest note: this pack only loads the 150-second variants, not the newer 240-second (4-minute) SongBloom weights that landed later - so if you've already grabbed the bigger model from HuggingFace, it won't appear in the dropdown.