Anima (Cosmos-Predict2) LoRA/LoKr trainer for Windows. 中文说明请见: README.zh-CN.md

• Trainer entry: anima_train.py (orchestration) + utils/* modules
• Config via TOML: --config <path-to-toml>
• Supports lora and lokr (LyCORIS full mode for LoKr)
• Optional ComfyUI key conversion on save
• Optional W&B logging
• LR schedulers follow HF warmup/decay behavior (constant, constant_with_warmup, linear, cosine)
• Supports radam_schedulefree optimizer

• anima_train.py: main training script
• utils/config_loader.py: TOML config loader
• config/save/anima_lora_config.example.toml: generic config template
• start_webui.bat: build frontend and launch WebUI backend
• install_dependencies.bat: one-click dependency installer (Windows)
• modelling/: modeling code
• anima/text_encoders/: tokenizer/config assets

Two ways to prepare config, one unified way to train:

1. Open WebUI or edit TOML directly.
2. Run training from terminal.
3. TensorBoard is enabled automatically by default.

1. Install dependencies:

install_dependencies.bat

2. Start WebUI:

start_webui.bat

3. In WebUI, edit and save config (config/save/*.toml).
4. Save config only. WebUI is a config editor; training is started from CLI.

Notes:

• WebUI only edits/validates config and does semantic checks in real time.
• Invalid configs (for example radam_schedulefree + cosine) are blocked from save.

1. Install dependencies:

install_dependencies.bat

2. Create a local config:

copy config\save\anima_lora_config.example.toml config\save\my_train.toml

3. Edit config\save\my_train.toml (model path, data path, output fields).
4. Run training:

.venv\Scripts\python.exe anima_train.py --config .\config\save\my_train.toml

Notes:

• TensorBoard event writing is on by default (--tensorboard-enabled).
• CLI training auto-starts TensorBoard by default (--tensorboard-autostart) on port 6006 (auto-increments if occupied).
• Disable auto-start with --no-tensorboard-autostart.

Weight files are not stored in this repository. Place your local model files under paths you configure, for example:

• anima/diffusion_models/*.safetensors
• anima/vae/*.safetensors
• anima/text_encoders/ (tokenizer + encoder files)

• Local configs are usually kept under config/save/*.toml.
• Use config/save/anima_lora_config.example.toml as the shared template.
• For adamw8bit, install bitsandbytes separately if needed.
• radam_schedulefree requires lr_scheduler=constant and no warmup (lr_warmup_steps=0, lr_warmup_ratio=0).

• FHfanshu
• OpenAI Codex

• Community references for LoRA/LoKr training workflows
• This training script references community-circulated implementations (including upper directory v1.01). If any content infringes your rights, please contact for removal.
• PyTorch, Transformers, LyCORIS, Weights & Biases

MIT License for this repository code. Please follow original licenses/terms for upstream models and third-party assets.