YUE_Stage_B_Sampler

Four nodes into the YuE graph, this is where you actually hear something. YUE_Stage_B_Sampler is the finish line: it runs the Stage B upsampler over your scratch track, decodes it through the xcodec vocoder, mixes the vocal and instrumental stems, and hands you an audio waveform plus a path to a saved file. If you've been waiting to hear your lyrics sung, this is the moment.

How it works

Stage A's output (stage1_set) holds the interleaved vocal/instrumental codec tokens. Stage B - loaded by YUE_Stage_B_Loader - upsamples those into a much richer representation, then this node decodes each stem through the xcodec vocoder decoders. That's what vocal_decoder_ckpt and inst_decoder_ckpt are for: the decoder_131000.pth and decoder_151000.pth files from ComfyUI/models/yue. The node first writes a 16kHz reconstruction and mixes the stems, then runs the upsampler vocoder for a 44.1kHz mix, and finishes with a low-frequency energy-match pass to clean up the seams. That final step is why rescale exists - leave it on unless you have a reason not to.

Inputs that matter

• stage1_set - from YUE_Stage_A_Sampler's stage1_set output. Nothing happens without it.
• model - from YUE_Stage_B_Loader.
• vocal_decoder_ckpt / inst_decoder_ckpt - dropdowns that only list models/yue files containing "131" / "151". If they show only "none", the decoder files aren't where the code looks and the vocoder will fail.
• rescale - on by default; normalizes loudness on the final mix.

Outputs

• audio (AUDIO) - a 44.1kHz waveform you can wire straight into a PreviewAudio node or any audio playback node. The bundled example workflow feeds exactly this, and it's the fun part: hit run and your song comes out the other end.
• string - the full path to the final saved mp3 in ComfyUI/output, for when you want the actual file.

Install and models

Same shared pack install as the rest of the chain - clone, requirements, restart:

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

(ComfyUI Manager → "ComfyUI_YuE" also works.) This node is the reason decoder_131000.pth and decoder_151000.pth must be in ComfyUI/models/yue - without them there's no vocoder. ckpt_00360000.pth and the semantic model are needed too, and the Stage B repo downloads itself on first load. Total model footprint is roughly 15GB, so budget disk before you start.

Gotchas

• Decoder dropdowns stuck on "none". It's the "131"/"151" filename filter again - put both .pth files in models/yue before you go looking for them.
• OOM in Stage B. If the upsampler runs out of VRAM, lower stage2_batch_size on the Stage B Loader rather than fighting this node.
• Missing stage1 files. stage1_set is a real object wire, but the underlying tokens live on disk. If you re-run this node against a stale graph, make sure YUE_Stage_A_Sampler actually ran in this session or the .npy files won't exist.
• It's slow, and that's normal. Stage B is a second full model pass, and the mix/vocoder steps add real time on top of Stage A's. A 30-second song on a modest GPU is a minutes-long operation; don't mistake it for a hang.