None Trainer

Choose based on your CUDA version:

# CUDA 12.8 (RTX 40 series recommended)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128

# CUDA 12.4
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

# CUDA 12.1
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

# CUDA 11.8 (older GPUs)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

Flash Attention significantly reduces VRAM usage and speeds up training.

Linux - Download from Flash Attention Releases:

# Check your environment versions
python --version                                      # e.g.: Python 3.12
python -c "import torch; print(torch.version.cuda)"  # e.g.: 12.8

# Download matching version (example: Python 3.12 + CUDA 12 + PyTorch 2.5)
wget https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.3/flash_attn-2.7.3+cu12torch2.5cxx11abiFALSE-cp312-cp312-linux_x86_64.whl

# Install
pip install flash_attn-*.whl

Windows - Download prebuilt from AI-windows-whl:

:: Example: Python 3.12 + CUDA 12.8 + PyTorch 2.9.1
pip install https://huggingface.co/Wildminder/AI-windows-whl/resolve/main/flash_attn-2.8.3+cu128torch2.9.1cxx11abiTRUE-cp313-cp313-win_amd64.whl

Tip: If no matching version exists, skip this step. The program will automatically use SDPA as fallback.

⚠️ Note: This project requires diffusers 0.36+ (dev version), install from git:

pip install git+https://github.com/huggingface/diffusers.git

# Clone project
git clone https://github.com/None9527/None_Z-image-Turbo_trainer.git
cd None_Z-image-Turbo_trainer

# One-click install dependencies
chmod +x setup.sh
./setup.sh

# Edit config (set model paths)
cp env.example .env
nano .env

# Start service
./start.sh

:: Clone project
git clone https://github.com/None9527/None_Z-image-Turbo_trainer.git
cd None_Z-image-Turbo_trainer

:: One-click install (double-click or command line)
setup.bat

:: Edit config (set model paths)
copy env.example .env
notepad .env

:: Start service
start.bat

After deployment, open browser: http://localhost:9198

# 1. Install Python dependencies
pip install -r requirements.txt

# 2. Install latest diffusers
pip install git+https://github.com/huggingface/diffusers.git

# 3. Install this project
pip install -e .

# 4. Build frontend (Important!)
cd webui-vue
npm install          # or pnpm install
npm run build        # generates dist directory
cd ..

# 5. Create config file
cp env.example .env

# 6. Start service
cd webui-vue/api && python main.py --port 9198

Besides Web UI, you can use command line directly:

# Generate Latent cache (VAE encoding)
python -m zimage_trainer.cache_latents \
    --model_path ./zimage_models \
    --dataset_path ./datasets/your_dataset \
    --output_dir ./datasets/your_dataset

# Generate Text cache (text encoding)
python -m zimage_trainer.cache_text_encoder \
    --text_encoder ./zimage_models/text_encoder \
    --input_dir ./datasets/your_dataset \
    --output_dir ./datasets/your_dataset \
    --max_length 512  # Optional: 256/512/1024, default 512

First copy example config and modify paths:

# Z-Image training
cp config/acrf_config.toml config/my_zimage_config.toml
# Edit my_zimage_config.toml, modify [model].dit and [[dataset.sources]].cache_directory

# LongCat-Image training
cp config/longcat_turbo_config.toml config/my_longcat_config.toml
# Edit my_longcat_config.toml, modify [model].dit and [[dataset.sources]].cache_directory

Then start training:

# Z-Image training (recommend using accelerate)
python -m accelerate.commands.launch --mixed_precision bf16 \
    scripts/train_zimage_v2.py --config config/my_zimage_config.toml

# LongCat-Image training
python -m accelerate.commands.launch --mixed_precision bf16 \
    scripts/train_longcat.py --config config/my_longcat_config.toml

⚠️ Important: Must modify these paths in config:

• [model].dit - Transformer model path
• [model].output_dir - Output directory
• [[dataset.sources]].cache_directory - Dataset cache path

# Load LoRA and generate image
python -m zimage_trainer.inference \
    --model_path ./zimage_models \
    --lora_path ./output/your_lora.safetensors \
    --prompt "your prompt here" \
    --output_path ./output/generated.png \
    --num_inference_steps 10

# Method 1: Use script
./start.sh          # Linux/Mac
start.bat           # Windows

# Method 2: Direct start
cd webui-vue/api
python main.py --port 9198 --host 0.0.0.0

# Method 3: Use uvicorn (hot reload)
cd webui-vue/api
uvicorn main:app --port 9198 --reload

# Service config
TRAINER_PORT=9198           # Web UI port
TRAINER_HOST=0.0.0.0        # Listen address

# Model paths
MODEL_PATH=/./zimage_models

# Dataset path
DATASET_PATH=./datasets

# Ollama config
OLLAMA_HOST=http://127.0.0.1:11434

[acrf]
turbo_steps = 10        # Anchor count (inference steps)
shift = 3.0             # Z-Image official value
jitter_scale = 0.02     # Anchor jitter

[lora]
network_dim = 16        # LoRA rank
network_alpha = 16      # LoRA alpha

[training]
learning_rate = 1e-4    # Learning rate
num_train_epochs = 10   # Training epochs
snr_gamma = 5.0         # Min-SNR weighting
loss_mode = "standard"  # Loss mode (see below)

[dataset]
batch_size = 1
enable_bucket = true
max_sequence_length = 512  # Text sequence length (must match cache)

💡 Beginners: Start with standard mode, try others if unsatisfied.

Scenario Configs:

Scenario Configs:

⚠️ Note: When both Freq and Style are enabled, high-freq penalties overlap (alpha_hf and lambda_tex). Consider reducing one.

• 📧 lihaonan1082@gmail.com
• 📮 592532681@qq.com

Apache 2.0

• Z-Image - Base model
• diffusers - Training framework
• Flash Attention - Efficient attention