Load VQGAN

This one's a deep cut, and it's worth saying upfront: it's the least documented corner of an already sprawling pack. Load VQGAN belongs to Mixlab's Fish Speech integration - Fish Speech (now rebranded OpenAudio) is a real, actively-discussed open-source voice-cloning TTS project - but the pack's own README, unlike almost everywhere else it lists a model, never spells out where the Fish Speech checkpoint actually goes. If you land here from a search, know that you're in genuinely under-documented territory and budget time to dig into the upstream Fish Speech repo for the model paths this node needs.

Here's the mechanism, because it explains why this node exists and looks the way it does. Fish Speech generates audio in two stages: first a text-to-semantic step (Mixlab's Prompt2Semantic node handles that half) turns your input text into a sequence of discrete audio tokens, then a vocoder decodes those tokens back into an actual waveform. VQGAN - a vector-quantized GAN - is that vocoder. Load VQGAN loads it and hands back a vqgan object that downstream Fish Speech nodes use to turn Prompt2Semantic's token codes into sound you can actually hear.

The node has three settings. config is a dropdown that, per the node's own schema, currently has exactly one option: firefly_gan_vq.yaml - "Firefly" being Fish Speech's name for this vocoder architecture, so there's nothing to actually choose here yet, just a fixed default. model is also a dropdown, but its schema reports zero available choices out of the box - meaning it scans a models folder and shows you whatever checkpoint files you've placed there, and until you've downloaded and dropped in a Fish Speech VQGAN checkpoint, that dropdown will simply be empty. device lets you pick cuda or cpu, with cuda as the sane default for anything beyond a quick sanity check.

Whether it's actually worth chasing down that model is a fair question. Community sentiment on Fish Speech's voice cloning is mixed rather than glowing - people who've used it report the output quality "not as good as it should be, not robust as well" compared to competing TTS options, and at least one recent report described locking the seed and still not getting reproducible output between runs on essentially the same input. It's a legitimate open project worth trying, just go in without assuming it'll be flawless the first time.

Install is the standard path for the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/shadowcz007/comfyui-mixlab-nodes.git
cd comfyui-mixlab-nodes
install.bat

or via ComfyUI Manager (search "comfyui-mixlab-nodes"), or pip3 install -r requirements.txt in a venv. That gets you the node itself. Getting it functional is a separate step the README doesn't walk you through here - you'll need to track down a compatible Fish Speech VQGAN checkpoint from the upstream fishaudio/fish-speech project and place it wherever Mixlab's model-scan expects it, which in practice means checking the pack's GitHub issues or Discord for the exact folder path, since it isn't in the README section covering the rest of the pack's model downloads. If model shows an empty dropdown after install, that's expected - it means no checkpoint, not a broken install.