IndexTTS Unload（释放显存）

One input, no outputs, one job: kick the IndexTTS2 model out of VRAM. It's the smallest node in the pack and the one you'll be glad exists. TTS is the cheap part of any audio pipeline - the expensive compute is always the video the voice drives - and IndexTTS2 is an 11.8GB model that defaults to FP32. If you generate a voiceover and then run a Wan or LTX render in the same ComfyUI session, that model is sitting in memory eating the card your video needs.

What it does

Give it the INDEXTTS_MODEL handle from IndexTTSLoader (the same one your generation node consumed), and it tears down the cached model: it offloads everything in the pack's model cache, runs a garbage collection, and calls torch.cuda.empty_cache(). The VRAM comes back for the next stage of the pipeline. There's nothing to configure - the input is required, so this node is a deliberate act, not something you sprinkle in by accident.

Where to put it

The intended shape of a real workflow:

IndexTTSLoader → IndexTTSSingle (or IndexTTSBatch) → ... → IndexTTSUnload → [video nodes]

Run your TTS pass, save the audio, evict the model, then let the video stage have the whole card. It's marked as an output node in ComfyUI, so it works as the terminal end of a TTS branch even though it produces nothing - which is exactly what a "free VRAM before continuing" node should be. Wire it up after the last generation node, not in the middle of your voice chain.

One honest caveat: it clears this pack's model cache. It won't unload your checkpoint, your VAE, or anything another custom node cached. If you want a wholesale nuke you'll want a generic free-VRAM node too - but for the specific job of "I'm done with TTS, give the GPU back," this is the one that's guaranteed to get IndexTTS2 out.

Installing it

Same as the rest of the pack - it ships together, so there's nothing extra to fetch. ComfyUI Manager: search IndexTTS2-PauseControl (it's on the Registry; if a brand-new node hasn't shown up in Manager's search yet, git clone https://github.com/lynx-gt/IndexTTS2-PauseControl into custom_nodes/, run python install.py, restart). The model weights are the big download, and they're not in the repo: ~11.8GB from the official index-tts release into ComfyUI/models/index_tts/.

Why you'll actually use it

Because the alternative is an OOM halfway through a video render, or babysitting --lowvram mode and eating a slowdown you didn't need. The whole audio-in-ComfyUI layer is bolt-on tech with dependency friction, but this node is one of the rare pieces of plumbing that just works: put it at the seam between the voice stage and the video stage, and the 11.8GB problem quietly goes away.