SeedVC4SingNode

What it's for

SeedVCNode (the speech node in the same pack) converts speech. SeedVC4SingNode exists for the case that node can't touch: what if your source is a song? Singing is voice conversion's hard mode - you can't just swap the timbre and drop the melody, or you get a flat monotone wearing a famous voice. This node keeps the tune while changing the singer, which is precisely the AI-cover pipeline people actually run.

It's the same zero-shot Seed-VC family and the same author (AIFSH), but it loads the fine-tuned 44 kHz singing model - the one the README calls out as a large quality jump for high-pitched singing - rendering with BigVGAN at 44.1 kHz instead of the speech node's 22.05 kHz. For covers, that resolution difference is audible.

How it's different under the hood

Same Whisper content encoder, same CAMPPlus speaker embedding, same diffusion transformer. The pitch handling is what changes:

• An RMVPE F0 extractor tracks the melody of both the source vocal and the reference singer. (RMVPE is the same pitch model RVC ships, if you've been around that ecosystem.)
• auto_f0_adjust (on by default) shifts the converted vocal's pitch contour onto the target singer's median register - the code literally aligns the two medians - so the result sits in a range that voice can plausibly sing.
• pitch_shift then nudges it an explicit number of semitones on top, by multiplying the F0 by 2^(n/12).

The inputs that matter

• source / target - same as the speech node: the vocal you're converting, and the singer voice you want. The reference is trimmed to ~25 seconds internally.
• auto_f0_adjust - the toggle above. Leave it on unless the original singer already sits in the target's register and you want the raw contour preserved.
• pitch_shift - semitones, default 0. Dropping a song a couple semitones down is the usual move so it fits the target singer's range.
• diffusion_steps - same story as its sibling: 10 by default, 50–100 for keepable quality.
• length_adjust / inference_cfg_rate - same meaning as the speech node; leave the CFG rate alone.

Output is a single AUDIO, ready for PreviewAudio, SaveAudio, or mixing back against the instrumental.

Install & gotchas

Same pack, same install: ComfyUI Manager → search "SeedVC-ComfyUI", or

cd ComfyUI/custom_nodes
git clone https://github.com/AIFSH/SeedVC-ComfyUI
# restart ComfyUI

First use triggers two extra downloads on top of what the speech node already grabbed - the RMVPE pitch model and the 44 kHz BigVGAN, both into ComfyUI/models/AIFSH/SeedVC/. First run is a big download either way, and the same dependency caveats apply: the pack pins onnxruntime-gpu even though the code only uses it on a Windows DirectML path, and the git-installed whisper package expects ffmpeg on your PATH.

Where people get burned:

• Feed it the isolated vocal stem, not the full mix. Convert the vocal-split track, then re-mix. This is a voice converter, not a remixer.
• The sing node always loads the fine-tuned 44 kHz model and the speech node never does - they don't share weights - so your only real quality dials are diffusion steps and the two pitch toggles.
• Long songs are chunked into 30-second windows with crossfaded overlap, so they work, but every chunk runs full diffusion. A whole track at 100 steps is minutes of GPU time; dial in settings on a short section first.