IndexTTS2-音色参考+情感描述

The friendliest of the pack's four IndexTTS2 variants: LamIndexTTS2Node3 (音色参考+情感描述) clones a voice from a reference clip, then lets you describe the emotion in words instead of hunting for the perfect mood reference or fiddling with eight sliders. Something like emo_text: "whispering, a little nervous" and the delivery follows.

Same architecture as its siblings: it doesn't run IndexTTS2 inside ComfyUI. It's a client for a separate IndexTTS2 HTTP service on http://localhost:5000 - the node checks the model is loaded, submits the job to /generate, polls until it's done, and downloads the wav into your output directory. IndexTTS2 itself is the open-source IndexTeam model whose emotion control is one of the reasons it caught on in the ComfyUI TTS crowd.

Inputs

• audio - the timbre reference (a file path or an AUDIO tensor); this sets the voice. The emotion comes from emo_text, not from this clip.
• text - the words to synthesize. Multiline, so paragraphs are fine.
• emo_text - the emotion description. The author's tooltip is just "情感描述文本"; the server interprets it however your IndexTTS2 build does, so natural-language moods work best ("excited", "sad, barely holding it together").
• server - http://localhost:5000, tooltip 请勿修改 ("don't modify").
• params (optional) - LamIndexTTS2AdvancedParams for sampling overrides.
• audio_dir (optional) - subfolder under ComfyUI's output directory for the result.

Outputs

• audio_path - full path to the generated audio.
• task_id - the job id on the TTS server.

Install

Part of the ComfyUI_Lam pack:

cd ComfyUI/custom_nodes
git clone https://github.com/yanlang0123/ComfyUI_Lam

or install via ComfyUI Manager and run the README's install + 修改文件 scripts. Grab the IndexTTS-2 model with the pack's 下载模型.bat (it pulls IndexTeam/IndexTTS-2 from Hugging Face). Then start the separate IndexTTS2 server on port 5000 - that's the step that trips everyone up, because nothing in ComfyUI tells you the model lives in a different process.

Gotchas

Standard family failure mode: no service on port 5000 means the node throws a connection error. Beyond that, the emotion description is only as good as what your server maps it to - vague emo_text like "good" gets you vague emotion, so be specific. And if your reference clip already carries a strong mood and you'd rather it just copy that, you don't need this node at all - that's LamIndexTTS2Node0 with one less input to fill.