dots.tts Edit 模型加载器 · T8star-Aix

T8_DotsTTS_EditModelLoader is the twin of the regular model loader, with one crucial difference: it loads dots.tts.edit, the dedicated speech editing checkpoint - not a TTS checkpoint. If you're building a workflow that rewrites or fixes words inside an existing recording (via T8_DotsTTS_Edit), this is the loader you wire in. Grab the wrong loader and the Edit node will refuse the handle with a "please select a dots.tts.edit model" error, which is the pack's way of keeping you from mixing the two runtimes.

Mechanically it's the same discipline as T8_DotsTTS_ModelLoader: the model_name dropdown is populated from a pinned catalog of known models (default dots.tts Edit · Speech Editing [edit]), it resolves your local copy from the standard ComfyUI/models/TTS/dots.tts/dots.tts.edit/ layout (plus extra_model_paths.yaml TTS paths and the legacy ComfyUI/models/dotstts/ folder), validates file sizes, and hands you a T8_DOTS_TTS_EDIT_MODEL handle. The catalog pins the exact Hugging Face repo and revision, and verify_hashes (off by default, since it re-reads ~5 GB) turns on full SHA-256 checking. Weights don't actually load until your first edit - the node description says "lazy load," and it means it.

The inputs you'll actually touch:

• precision - bfloat16 default, float32 for odd hardware. Leave it.
• optimize - torch.compile, off by default. First compile is slow and can fail; the tooltip is explicit: "disable it when it fails." This one's worth trying on long projects but not on a quick test.
• release_after_run - free the weights after each edit run. Turn it on if you're swapping models a lot and VRAM is tight.
• custom_model_path - an absolute path override if your checkpoint lives somewhere non-standard.

The outputs are the model handle (into T8_DotsTTS_Edit) and a model_info JSON with repo id, pinned revision, directory and precision - useful for confirming you're running the exact revision the pack was tested against.

The gotcha that bites everyone

The edit model is a separate ~5 GB download. It is not bundled, and the regular download_models.py --model soar command won't get it. You need:

python scripts\download_models.py --model edit --comfyui-root D:\ComfyUI

First-timers load the TTS model fine, wire up an Edit node, and hit "model missing" because the edit checkpoint was never fetched. The loader won't download anything on import by design - that's a deliberate, security-conscious choice (the pack follows the ComfyUI Registry no-auto-install norm), so the download step is always on you.

Otherwise installation is the standard pack install - ComfyUI Manager search dots-tts-t8 or git clone https://github.com/T8mars/comfyui-dots-tts-T8 into custom_nodes plus pip install -r requirements.txt. Python 3.10–3.12 and PyTorch/torchaudio 2.8+ with matching minor versions are the environment requirements, same as the rest of the pack.