YuE2 VAE 解码

Last stage of the advanced chain, and the shortest to explain: this is the VAE. Latents from YuE2 声学合成 go in, a 48 kHz stereo waveform comes out.

What happens inside

The worker loads the latent array, checks it against the manifest chain, then runs the VAE decoder with halo-cropped chunking - 1024 frames at a time by default, dropping to 512 when your memory budget is 12 GiB or less. Halo cropping means each chunk carries the overlap it needs from its neighbours, so you don't get seams at the chunk boundaries; the audio is decoded as one continuous thing, just not all at once. The artifact it writes is audio.flac at 48 kHz, 24-bit PCM.

Two required inputs - model and latents - and no settings, deliberately. Everything that governs this stage (budget, chunking) is decided upstream and travels in the handle.

Outputs

audio - an AUDIO object for Preview Audio or Save Audio. result - a YUE2_RESULT for YuE2 导出工件, which is how you keep the artifact bundle rather than just the FLAC. metadata - the job status JSON, with sample rate, duration and timings.

The check that will annoy you exactly once

Decode refuses any latent file that isn't the canonical one recorded in its manifest. This is literal: the worker compares your path against <artifact-dir>/latent.npy and raises if it differs. So you can't point this node at a hand-copied latent array, a re-saved version, or a file you moved into a tidy folder of your own. It also verifies the semantic and plan manifests underneath it before decoding.

That's not paranoia for its own sake - it's the difference between "this audio came from those tokens from that plan from that model" and "here's some noise I found". Advanced-stage workflows are the ones where people later can't reproduce anything, and this is the pack refusing to help you get there.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-YuE2-T8.git

Run install_runtime.bat once from the node directory, restart ComfyUI, and let YuE2 模型服务 pass its check before queueing. Registry: comfy node install yue2-t8. Windows 10/11, NVIDIA GPU, 24 GB VRAM and 60 GB free disk recommended.

Where people get burned

Cancellation lags here, specifically. The pack's own documentation warns that cancellation may be delayed during VAE decode, and it's the stage where you'll notice it most - the worker finishes the current chunk before it checks. If you force-kill the worker anyway (force_cancel on YuE2 卸载/取消), the job dies, and the next job pays for a cold start: the model has to reload into the service before anything happens. Sometimes that's the right trade; just know you're trading one slow minute for another.

A truncated song decodes to a truncated song. If the semantic stage flagged truncation, the audio you get back will end early or lose its final section. The decode stage faithfully reproduces a problem that happened three nodes earlier - go back and shorten the source lyrics.

Audio looks shorter than you expected. Compare metadata against what you asked for. Songs are bounded by the token budget in the plan, not by your hopes.

Nothing in the graph shows progress. For a long decode, the WebUI at 127.0.0.1:8189 shows the job's stage and per-stage memory; the ComfyUI console is your fallback.