Easy SongGeneration - 下载模型

The Easy SongGeneration - 下载模型 node - "Download Model" - is where every workflow in this pack starts, because the pack ships no weights at all. Tencent's SongGeneration music model is a multi-gigabyte download, and this node is the only sane way to land it in the right folder without hunting through Hugging Face tree views. It pulls common, third_party (the Demucs separator), the auto-reference-style weights, and the checkpoint you pick straight into ComfyUI/models/SongGeneration/ - the exact layout the loader node expects.

If you've done this dance with any other audio pack, you know the drill: audio in ComfyUI is a bolt-on layer where the models are real but the integration is a maintained workaround. This node is that workaround done properly - one run, everything in place, nothing left to a manual git lfs pull.

What it actually downloads

The node calls huggingface_hub.snapshot_download against eastmoe/SongGeneration - the pack author's mirror of Tencent's weights - and writes into ComfyUI/models/SongGeneration/. Three things always come down regardless of your 模型 choice:

• common/ - the audio tokenizer, VAE and codec weights every checkpoint shares.
• third_party/demucs/ - needed only if you feed the generation nodes a reference audio file.
• The auto-reference-style prompt weights, fetched from Tencent's Hugging Face Space.

The 模型目录 (model) dropdown then decides which checkpoint joins them: SongGeneration-v2-large (the big 4B one, ~22-28GB VRAM per the upstream table), SongGeneration-base-full, SongGeneration-base-new, runtime-only (just the shared assets), or all.

The inputs that matter

• 下载源 (source) - hf-mirror.com or huggingface.co. This is a pack by a Chinese author who knows that huggingface.co is a pain on mainland networks; the mirror is a Chinese CDN for HF. If you're outside China, huggingface.co is usually fine. If a download stalls, flip this and rerun.
• 模型目录 (model) - the checkpoint, above.
• 覆盖已有文件 (overwrite_existing) - default off, which skips any local file that already matches by size. Leave it off unless you're sure a download got corrupted.
• 下载线程 (download_threads) - concurrency for the snapshot, default 8. Crank it to 32 on a fat pipe.
• 关闭 SSL 验证 (disable_ssl_verify) - only for the rare case where a proxy or broken cert chain blocks HTTPS. You'll know when you need it.

It's an OUTPUT_NODE that reports a status string and fires every time you hit Run (IS_CHANGED returns NaN), but with overwrite off it's just a cheap consistency check on reruns.

Gotchas

The optional auto-reference-style weight comes from a Hugging Face Space, and some mirrors redirect it to an unreachable huggingface.co URL. That failure is non-fatal - the main download proceeds and text-to-song still works; only the Pop/Auto style presets on the generate nodes lose their reference audio. Rerun the node later or drop the file manually if you want those back.

After the download finishes, reopen the dropdown on Easy SongGeneration - 加载模型 - model choices are scanned at UI load, so it won't see the new folder until you refresh. And if you ever see a "Git LFS pointer, not model weights" error, it means something got downloaded as a pointer file rather than real weights - run this node with overwrite on rather than fighting it manually.