Anima TrainFlow

Anima TrainFlow is a streamlined, one-page GUI for training LoRA on the Anima 2B model. Optimized to run on hardware with as little as 6GB of VRAM, it eliminates technical overhead by focusing on the essential settings that impact training results the most.

⚡ Everything you need in one tool

A complete, ready-to-use workflow from start to finish.

Raw Images ➔ Smart Resize & Crop* ➔ Auto-Tag ➔ One-Click Train** ➔ Ready LoRA

• Smart Prep: In most cases, it only performs resizing. Intelligent cropping is triggered only for extreme aspect ratios that don't fit into training buckets.
• Fast Results: Average training time is ~1 hour (benchmarked on RTX 3060 12GB).

Quick Start (Portable)

1. Download Portable Version (3GB)
2. Extract the archive using 7-Zip or WinRAR.
3. Run start_trainer.bat.
4. Open the 🔧 Paths to Models <- Set Once accordion and specify the paths to your model files.
5. Specify your Dataset Path (images + .txt files) and Trigger Word, then click Start.

Manual Installation

If you prefer to set up the environment manually instead of using the portable version, follow these steps:

1. Clone the repository:

   git clone https://github.com/ThetaCursed/Anima-TrainFlow
   cd Anima-TrainFlow

2. Install dependencies:
   Run Install_Requirements.bat

3. Download Required Models: Run the following commands from the root folder:

   • WD Tagger (used for auto-captioning):

     git clone https://huggingface.co/SmilingWolf/wd-eva02-large-tagger-v3 models/wd-eva02-large-tagger-v3

   • U2Net Model (used for Smart Cropping):

     curl -L -o models/u2net/u2net.onnx https://github.com/danielgatis/rembg/releases/download/v0.0.0/u2net.onnx

4. Launch the Trainer: Run start_trainer.bat

Key Features

• Zero-Tab Interface: All critical parameters (Trigger Word, Rank, LR, Steps) are accessible on a single screen.
• Live Training Previews: Watch your LoRA improve in real-time. The built-in gallery automatically updates whenever a new sample is generated.
• AI-Powered Smart Cropping: Integrated U2Net model automatically performs subject-aware, head-first cropping and resizes images to optimal aspect-ratio buckets via multi-threading.
• Built-in Auto-Captioning: Integrated WD14 Tagger (EVA02 v3) automatically generates multi-threaded .txt tags for your dataset with customizable general and character thresholds.
• Pre-Flight Validation: Automatically scans the dataset for missing captions, oversized images (>=2048px), and missing model paths to prevent crashes before training starts.
• Persistent Sessions: All UI inputs, paths, and slider positions are instantly auto-saved and restored on the next launch.
• Portable Edition: Includes an embedded Python environment to avoid installation or complex setup.
• Low VRAM Friendly: Specifically tuned for 6GB+ NVIDIA GPUs with aggressive RAM/VRAM clearing between tasks.
• Optimized Defaults: Pre-configured for BF16 precision and latent caching to ensure maximum performance and stability.
• Prodigy Native: Intelligent Learning Rate handling and optimized defaults for the Prodigy optimizer.

Dataset Preparation

Place all your training images (.png, .jpg, .webp) in a single folder. Every image must have a matching text file with the same name containing its tags/captions (e.g., image1.png and image1.txt). You can easily generate these text files using the built-in Auto-Caption Dataset tool.

System Requirements

• OS: Windows 10/11.
• GPU: NVIDIA GPU (6GB+ VRAM recommended for Anima 2B training).
• Storage: ~6.5GB of free space (SSD recommended).

Technical Details

• Core: Based on a modified version of sd-scripts for Anima 2B architecture.
• UI: Built with Gradio featuring a customized dark theme.
• Backend: Utilizes accelerate launch for optimized execution.
• Auto-Save: All paths and configurations are automatically saved to settings.json.