INDEX-TTS2 折腾记录

前言

之前折腾了MOSS-TTS,看到对比INDEX-TTS效果更好,再折腾一下,记录一下。

直接运行

uv venv
source .venv/bin/activate
uv sync --all-extras

结果报错了

Resolved 185 packages in 10.23s
error: Distribution `triton-windows==3.1.0.post17 @ registry+https://mirrors.sustech.edu.cn/pypi/web/simple` can't be installed because it doesn't have a
       source distribution or wheel for the current platform

hint: You're on Linux (`manylinux_2_41_x86_64`), but `triton-windows` (v3.1.0.post17) only has wheels for the following platform: `win_amd64`; consider adding "sys_platform == 'linux' and platform_machine == 'x86_64'" to `tool.uv.required-environments` to ensure uv resolves to a version with compatible wheels
uv sync

这次可以了。

uv tool install "huggingface-hub[cli,hf_xet]"
HF_ENDPOINT="https://hf-mirror.com" hf download IndexTeam/IndexTTS-2 --local-dir=checkpoints
uv run webui.py

结果报错了

ModuleNotFoundError: No module named 'gradio'

还没有安装 webui

uv sync --extra webui

装上再次运行

uv run webui.py

可以了,又自动下载了一些模型文件,最终还是炸显存了。

改用

uv run webui.py --fp16

起来了,打开 http://127.0.0.1:7860 上传reference音频,输入文字,第一次生成成功了,第二次又炸显存了。

通过降低gui_seg_tokens可能有点用,但是Text长度还是最长90个中文字左右。

uv run webui.py --fp16 --gui_seg_tokens 80

最后修改于 2026-07-22