English 中文（正體）

a CustomTkInter GUI for processing and training DiffSinger models

DiffTrainer brings together the most useful tools for DiffSinger in one easy, graphical package.

• nnsvs-db-converter for converting wav+lab data to wav/ds+csv
• SOME for estimating pitch
• DiffSinger's primary training
• OpenUtau export scripts

setup options

If you have an NVIDIA GPU:

• make sure a compatible version of CUDA Toolkit is installed
  • current compatible versions: 11.8, 12.1, 12.4, 12.6, 12.8, 12.9
• proceed to next 'if'

If you have never used Python:

• run conda_installer.bat
• "Update Tools" once DiffTrainer is running
• use run_gui.bat to launch after that

If you have used Python:

• DiffTrainer by default uses Miniconda to manage conflicting package requirements.
• To use an existing conda installation:
  • (Windows/Mac) install requirements.txt to base environment(3.10 strongly recommended, other versions may still work for the base environment)
  • (Linux) requirements.txt still works, but has graphical bugs(described below). a fixed environment only for launching is in the assets folder
  • run setup_conda_envs.py to configure the required environments
  • "Update Tools" on the first tab to complete the install
• As of v0.2.1, the names of the environments are hardcoded requirements.

Known issues:

• langloader editor usually hides behind main window
• ~~honestly langloader is just ugly and bad~~ improved as of 0.3.14(1/16/25)
• if you type in the save interval or batch size boxes, an error appears in the terminal window
  • no actual impact, just enter your number and ignore it
• ~~if you load one dataset, change your mind and load another, it fails to write the config~~ fixed in 0.3.29(5/7/25)
• do not name checkpoint folders just "acoustic" or "variance", it conflicts with the onnx export cleanup
• (Linux only) if the text and buttons appear jagged:
  • ~~in the base environment, conda install --channel=conda-forge tk[build=xft_*]~~ NO LONGER RECOMMENDED. this used to work but now renders the entire conda install non-functional.
  • improved 7/10/25, but not fully fixed. you can use assets/linuxbaseenv.yml to create a fixed environment, but then you have to launch from difftrainerBase
  • ex. conda env create -f assets/linuxbaseenv.yml
• (Intel Mac only) dependency of a dependency libcst no longer packaged for x86
  • add libcst to upper half of environmentA/B.yml before running setup_conda_envs.py(unsure if this will continue to work or what other impacts the outdated version has)

language support

DiffTrainer uses ez-localizr to allow GUI language selection. All users are welcome to translate the text found in en_US to other languages and submit a pull request.

to do

soon

• better readme

eventually

• ~~advanced export~~
• more translations
• ~~an icon that isn't amogus~~