A complete, self-contained, browser-based Windows LoRA trainer for the Anima diffusion architecture.
This project was built by taking the excellent core training scripts from Kohya-ss and heavily modifying them to support the bleeding-edge Anima architecture (Qwen3 Text Encoder + DiT + Custom VAE).
Instead of dealing with complex command lines, missing dependencies, and confusing python environments, this project packages everything into a standalone, portable bundle. You run the installer once, and it automatically sets up an isolated environment with its own Python and dependencies. Then, you manage all your training via a sleek, interactive Web UI right in your browser.
- Standalone Portability: Installs its own Miniconda and Python 3.13. Doesn't mess with your system's global Python.
- Web-based UI: Manage training jobs, configure datasets, edit TOML files, and tweak hyper-parameters without touching a single line of code.
- Live Image Generation & Sampling:
- Sample images are automatically generated during training at your specified intervals (e.g. every N epochs or steps).
- The UI has a dedicated Samples Gallery where you can view your generated outputs live as the model trains!
- Integrated TensorBoard: One-click TensorBoard launch to monitor your loss graphs directly in the browser. It automatically finds free ports and connects.
- Multi-GPU & Optimization: Supports DeepSpeed, Fully Sharded Data Parallel (FSDP), Tensor Parallelism, Flash Attention, and multiple optimizers including 8-bit AdamW and Prodigy.
- Hardware Monitoring: Live GPU VRAM, utilization metrics, and CPU usage are displayed directly in the header of the Web UI.
To train an Anima model successfully, you will need:
| Requirement | Specification |
|---|---|
| OS | Windows 10 / 11 (64-bit) |
| GPU | NVIDIA GPU with CUDA 12.8 support |
| VRAM (LoRA) | 8 GB minimum (12 GB+ recommended for higher batch sizes or higher ranks) |
| VRAM (Full Finetune) | 24 GB minimum (Full Checkpoint fine-tuning requires significantly more VRAM than LoRA) |
| RAM | 16 GB system RAM minimum |
| Disk Space | ~15 GB free for the isolated Python environment, PyTorch, and model weights |
| Node.js | v18 or later to run the Web UI |
If you ever move or rename the Anima-Potable-Trainer folder to a different drive or location after installation, you MUST double-click the SetEnv.bat file once.
This will automatically fix all the internal environment paths to match the new location and prevent critical errors during training.
🚨 IMPORTANT NOTICES 🚨
- The Anima Base Model: The Anima base model weights (DiT, Qwen3, VAE) are subject to strict licensing. They are for NON-COMMERCIAL USE ONLY. You may use them for personal projects, research, and non-commercial creative work. You must follow their official licensing. Do not use the base model or any LoRAs trained on it for commercial products, paid services, or profit without explicit permission from the original creators.
- This Software: This trainer UI, batch scripts, and wrappers are released under the MIT License. The underlying modified Kohya-ss scripts remain under the Apache 2.0 License.
- Download and extract this repository to a folder (e.g.
C:\AnimaTrainer).- Note: If you ever move this folder to a different drive or location later, you must double-click the
SetEnv.batfile once to update the paths!
- Note: If you ever move this folder to a different drive or location later, you must double-click the
- Double-click
download_models.batto automatically download the required base models.- It will prompt for your Hugging Face token (optional) and place the following 3 files in the
models/directory:anima_baseV10.safetensorsqwen_3_06b_base.safetensorsqwen_image_vae.safetensors
- It will prompt for your Hugging Face token (optional) and place the following 3 files in the
- Double-click
setup.batto install the isolated Python environment. - Double-click
training-ui\start_windows.batto launch the Web UI. - Open your browser to http://localhost:3000.
Anima uses a Diffusion Transformer (DiT) backbone with Qwen3-0.6B as the text encoder and a custom image VAE. It utilizes a flow-matching training objective with logit-normal timestep sampling.
- Kohya-ss — Core training library and LoRA infrastructure: https://github.com/kohya-ss/sd-scripts
- Anima — The diffusion model architecture.
- Qwen3 — Text encoder by Alibaba Cloud: https://huggingface.co/Qwen
- Hugging Face — Diffusers and Transformers libraries.