Top TTS 2.5 - Unload Model

Top TTS 2.5 - Unload Model is the boring node at the end of the workflow that saves your GPU when the fun is over. You feed it the model handle from the loader, it closes the model down, requests a VRAM cache cleanup, and returns a status string. That's it. It's an output node, so it sits at the end of the graph like a save node, even though it's not saving anything - it's freeing.

Why you need it at all

Here's the thing about this pack's Load Model node: it keeps the model resident and warmed between runs. That's great for iterating on takes - you don't re-pay the load cost every generation. But IndexTTS 2.5 is a serious chunk of VRAM, and it's running in a separate worker process that holds onto that memory even when your image or video nodes are screaming for it. ComfyUI does not hand GPU memory back just because you stopped touching a node. If you're chaining this into a bigger pipeline - say, TTS voice first, then a Wan or LTX video render, or even just a different TTS engine - that resident model is in the way.

So the practical pattern is: generate all your lines, then run the unload, then let the rest of the graph have the VRAM. It's the "explicit model unloading" the README advertises, and it's genuinely useful on cards where you're bumping the ceiling.

The one trap

Unload is permanent. Once you hit it, that model handle is dead - the worker process is gone, and wiring the same handle back into Synthesize raises a clear "This Top TTS model was unloaded. Run the model loader again." error. That's by design, but it bites people who build one big workflow with unload at the end and then re-run only the top half, expecting the audio to still flow. It won't. Re-running the graph re-executes the loader anyway, so in a normal full run it's a non-issue - just don't unload between generations you care about.

Install

Nothing special here. This node rides along with the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/whmc76/ComfyUI-Top-TTS.git
python -m pip install -r ComfyUI-Top-TTS/requirements.txt
python ComfyUI-Top-TTS/install.py

Plus the weight download (download_models.py --source huggingface --accept-license) if you haven't already - the loader will refuse to run without it.

Verdict

A three-line utility that you won't need for your first session and will appreciate by your third, once the novelty wears off and the VRAM math starts mattering. It's not glamorous; it's the difference between a workflow that politely leaves when it's done and one that squats on your GPU forever.