dots.tts 释放模型 · T8star-Aix

Somewhere between your third generation and your next video render, you'll hit it: "CUDA out of memory" on a workflow that shouldn't need that much VRAM, because the TTS model you loaded an hour ago is still sitting in memory. T8_DotsTTS_Release is the targeted fix - it frees the dots.tts models and CUDA cache that this pack holds, without nuking the whole ComfyUI process the way some aggressive "free everything" nodes do.

The mechanism is straightforward and honest about its scope. Optionally connect a model handle (the output of T8_DotsTTS_ModelLoader or T8_DotsTTS_EditModelLoader) and the node releases that specific model's weights; leave it disconnected and it clears whatever TTS/ASR models and cached buffers the pack is currently holding. It's the manual version of the release_after_run toggle you'll find on the loaders - that flag does the same cleanup automatically after each generation, and this node does it on demand, mid-workflow, wherever you choose to put it in the graph.

The output is release_result, a JSON string reporting what was actually freed. Read it once to confirm the model is gone from VRAM before you queue your video job; if the report shows nothing was held, you saved nothing - which is also useful information.

Where it fits in practice

• Model-swapping workflows. If you alternate between SOAR and MF 2-Steps (or between TTS and a video model), dropping a Release node at the end of the TTS branch keeps the other branch from inheriting your VRAM debt.
• Long batch runs. BatchQueue and Project hold the model for the whole run by design; a Release node after them means the next workflow starts clean.
• Debugging OOM. The fastest way to test whether a leftover TTS model is the culprit is to run this node and re-run your workflow. If it suddenly fits, you found it.

It's a thin node by design - one optional input, one JSON output - and it costs nothing to have lying around. No model download needed to run it (though without a connected handle there's nothing to free). Installation is the pack itself: ComfyUI Manager search dots-tts-t8, or git clone https://github.com/T8mars/comfyui-dots-tts-T8 into custom_nodes and pip install -r requirements.txt.