lora-trainerlisted

# Lora Trainer — Anima LoRA 训练编排 (v2.3, ported) > **Plugin path**: `skills/lora-trainer/SKILL.md` > **Upstream**: L5 application skill. Load `chenxin-core` (L4) first for VRAM/recipe context — > this skill targets Anima 1.0 (~2B Cosmos DiT) so VRAM/quant choices from > `hardware/8gb.json` directly affect training defaults. ## 1. 工具概览 | 维度 | 值 | |------|-----| | **工具** | [gazingstars123/Anima-Standalone-Trainer](https://github.com/gazingstars123/Anima-Standalone-Trainer) | | **路径** | `E:/Comfy/Anima-Standalone-Trainer` | | **venv** | `E:/Comfy/Anima-Standalone-Trainer/venv`（Python 3.12 + torch 2.7 + CUDA 12.8） | | **入口脚本** | `skills/lora-trainer/scripts/train-anima-standalone.sh` | | **训练入口** | `accelerate launch anima_train_network.py --config_file <toml>` | | **VRAM** | < 6GB（Anima 小模型 + fused QKV + TP/SP 优化） | | **ComfyUI 依赖** | **不需要**（独立 venv） | | **配置方式** | toml 文件（训练 + 数据集分离） | | **Web UI** | `training-ui/start_training_ui_anima.bat` → `http://localhost:3000` | ## 2. 触发词 ``` "训练 LoRA" / "训 LoRA" / "train LoRA" / "lora training" / "训 Anima LoRA" / "训角色" / "训场景" ``` ## 3. 必需模型 | 模型 | 路径 | 大小 | |------|------|------| | DiT | `E:/Comfy/comfyui-licyk-20260608/core/models/checkpoints/anima_baseV10.safetensors` | 4.18GB | | Qwen3 TE | `E:/Comfy/comfyui-licyk-20260608/core/models/text_encoders/qwen_3_06b_base.safetensors` | 1.19GB | | VAE | `E:/Comfy/comfyui-licyk-20260608/core/models/vae/qwen_image_vae.safetensors` | 254MB | ## 4. 入口命令 ```bash # 最小可用 bash skills/lora-