IndexTTS 2.5 模型加载器 · T8star-Aix

IndexTTS 2.5 is the ~5GB box of weights that everything else in this pack hangs off of, and this loader is the socket it plugs into. It's the first node in any IndexTTS 2.5 workflow, and it's also where most first installs die - not because the node is broken, but because the weights aren't where it's looking. Worth saying up front: there's no API key, no cloud call, no huggingface login required at runtime. Everything runs on your own GPU, and this node is just the ComfyUI-native way of saying "find this directory, verify it, and keep it warm."

What it actually does

The loader scans ComfyUI/models/TTS/IndexTTS-2.5 - the pack deliberately does not keep weights inside custom_nodes, so multiple workflows can share one copy instead of duplicating gigabytes. It also honors any TTS path you've declared in extra_model_paths.yaml. Against the pack's pinned manifest (manifests/model_2_5.json) it runs a file-size sanity check on every weight, and if you flip verify_hashes it'll read all ~5GB and compare full SHA-256 sums. That's a "once, after a weird download" button, not a daily thing.

Two details in the source are worth knowing because they bite people. First, the manifest is fixed to 2.5 - this loader will not silently fall back to IndexTTS 2.0 weights the way some older wrappers do. Second, weights load lazily at the first generation and then stay cached with a global thread lock, so concurrent queues on the same model get serialized instead of corrupting the upstream codec caches. Under 10GB of VRAM it automatically switches to a low-memory strategy (long text gets segmented, and the emotion model is freed before synthesis).

The inputs that matter

• model_name - the dropdown of detected models. If you see the [未找到] 请将模型放入 models/TTS/IndexTTS-2.5 placeholder, that's the "you haven't placed the weights yet" signal, not a bug.
• device (auto / CUDA / CPU) and precision (auto / bfloat16 / float32). Leave both on auto unless you have a reason; auto picks bfloat16 on a supporting GPU and floats otherwise.
• acceleration_mode - defaults to off, which is the zero-dependency, most-compatible setting. The fancier modes (auto_safe, bigvgan_cuda, torch_compile, gpt_accel, deepspeed) are all optional and none get installed for you. If a mode can't run, it degrades to off and the model_info output tells you why - check that string before blaming the model.
• release_after_run - unloads this model after each run. Great for tight VRAM, slower for repeated generations.
• custom_model_path - absolute path to a complete 2.5 directory if your setup lives outside the standard layout.

Outputs: IndexTTS 2.5 模型 (wire into Generate or the segmentation preview) and 模型信息, a status string that's genuinely worth opening when things behave oddly.

Getting the weights in

The README's flow, minus the guesswork:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-indextts25-t8.git comfyui-indextts25-T8
cd ComfyUI && python -m pip install -r custom_nodes/comfyui-indextts25-T8/requirements.txt

Then download the models - ModelScope is usually fastest from Asia, Hugging Face elsewhere:

python ComfyUI/custom_nodes/comfyui-indextts25-T8/scripts/download_models.py --source huggingface --accept-license

After the download finishes you have to restart ComfyUI for the dropdown to refresh. The requirements file deliberately excludes torch/torchaudio so it can't clobber your existing CUDA build - resist the urge to "fix" that by installing torch yourself.

Common issues

Model doesn't appear in the dropdown: weights in the wrong folder, or you haven't restarted since placing them. The pack's environment check script (scripts/check_environment.py) will tell you both whether the model dir is valid and what acceleration is actually available. And if you're on a bleeding-edge ComfyUI that shipped transformers 5.x - the pack needs >=4.52.1,<5, and transformers 5 broke its generation API, so either use ComfyUI's bundled 4.57.6 (fine, verified) or expect hard failures.