[IMPORTANT]

** Commit ad401a8 should help stabilize distributed training on Windows. If you're still facing issues and previous fixes did not help, try setting GLOO_SOCKET_IFNAME to different networking devices**

A lightweight, decoupled training environment for circlestone-labs' Anima model, currently support Lora training only. Windows and Linux support. Built upon sd-scripts implementation.

• Python 3.10+ (Python 3.12 recommended)
• Node.js (Required for the Web UI)
• CUDA fitting your system (CUDA 12.7+ recommended)

git clone https://github.com/gazingstars123/Anima-Standalone-Trainer.git
cd Anima-Standalone-Trainer

Run the provided setup script for your operating system:

Windows:

.\setup_env.bat

Linux:

./setup_env.sh

This will create a virtual environment (venv), install all Python dependencies (assuming you have met the prereqisites), and set up the Web UI.

This script will probably install Torch and Torchvision version below. Depends on your system, you may want to install another version of Pytorch with CUDA.

pip install torch==2.7.0 torchvision==0.22.0 --index-url https://download.pytorch.org/whl/cu128

To start the training server and open the web interface:

Windows:

.\training-ui\start_training_ui_anima.bat

Linux:

./training-ui/start_linux.sh

Once launched, open your browser to: http://localhost:3000

After launching the UI for the first time, you'll need to configure your model paths:

1. Click the ** Global Settings** (gear icon) in the bottom-left corner
2. Set the following paths:
   • DiT Model Path — Path to your Anima DiT safetensors file (e.g. C:\model\anima.safetensors)
   • VAE Path — Path to the VAE model (e.g. C:\model\qwen_image_vae.safetensors)
   • TE Path — Path to the CLIP text encoder (e.g. C:\model\text_encoders\qwen_3_06b_base.safetensors)
   • Venv Path - Path to your local venv, venv can be reused if you redownload the repo
3. Click Save

These paths are saved globally and shared across all training jobs.

v2.0.0. Linux support, Multi-GPU inference

v1.1.0. Improving caching and others I/O performance.

Tested on torch2.7+cu128 and torch2.10+cu130 with this fix applied on Windows when encountered libuv error.

Seems to works best with torch<=2.3 and cuda <= 12.4 without directly applying the fix.

*NEW*

Adding support for multi-gpu inference

To update, simply run this command

git pull

Some features and settings from sd-scripts may not be available or working properly at the momment.

Built and tested on Windows 11, RTX 5080 + RTX 3090, 96GB DDR5, Python 3.12.1, CUDA 13.1, Pytorch 2.10