Z-Image-Turbo-Quantized

Quantized weights for Z-Image-Turbo optimized for 8GB VRAM GPUs.

📦 Available Models

• z_image_turbo_scaled_fp8_e4m3fn.safetensors (6.17 GB) - FP8 E4M3FN quantized weights
• z_image_turbo_int8.safetensors (6.17 GB) - INT8 quantized weights

🚀 Installation

git clone https://github.com/ModelTC/LightX2V.git
cd LightX2V
pip install .

💻 Usage for 8GB VRAM GPUs

To run Z-Image-Turbo on 8GB VRAM GPUs, you need to:

1. Use quantized transformer weights (FP8 or INT8)
2. Use int4 quantized Qwen3 text encoder
3. Enable CPU offloading

Complete Example

from lightx2v import LightX2VPipeline

# Initialize pipeline
pipe = LightX2VPipeline(
    model_path="Tongyi-MAI/Z-Image-Turbo",
    model_cls="z_image",
    task="t2i",
)

# Step 1: Enable quantization (FP8 transformer + INT4 text encoder)
pipe.enable_quantize(
    dit_quantized=True,
    dit_quantized_ckpt="lightx2v/Z-Image-Turbo-Quantized/z_image_turbo_scaled_fp8_e4m3fn.safetensors",
    quant_scheme="fp8-sgl",
    # IMPORTANT: Use int4 Qwen3 for 8GB VRAM
    text_encoder_quantized=True,
    text_encoder_quantized_ckpt="JunHowie/Qwen3-4B-GPTQ-Int4",
    text_encoder_quant_scheme="int4"
)

# Step 2: Enable CPU offloading
pipe.enable_offload(
    cpu_offload=True,
    offload_granularity="model",  # Use "model" for maximum memory savings
)

# Step 3: Create generator
pipe.create_generator(
    attn_mode="flash_attn3",
    aspect_ratio="16:9",
    infer_steps=9,
    guidance_scale=1,
)

# Step 4: Generate image
pipe.generate(
    seed=42,
    prompt="A beautiful landscape with mountains and lakes, ultra HD, 4K",
    negative_prompt="",
    save_result_path="output.png",
)

⚙️ Configuration Options

Quantization Schemes

FP8 (Recommended) - Better quality and speed:

dit_quantized_ckpt="lightx2v/Z-Image-Turbo-Quantized/z_image_turbo_scaled_fp8_e4m3fn.safetensors",
quant_scheme="fp8-sgl",

INT8 - Alternative option:

dit_quantized_ckpt="lightx2v/Z-Image-Turbo-Quantized/z_image_turbo_int8.safetensors",
quant_scheme="int8-sgl",

Offload Granularity

• "model" (Recommended for 8GB): Offload entire model to CPU, load to GPU only during inference. Maximum memory savings.
• "block": Offload individual transformer blocks. More fine-grained control.

⚠️ Important Notes

1. Order matters: All enable_quantize() and enable_offload() calls must be made before create_generator(), otherwise they will not take effect.

2. Text encoder quantization: Using int4 Qwen3 text encoder is highly recommended for 8GB VRAM GPUs to ensure stable operation.

3. Memory optimization: The combination of FP8/INT8 transformer + int4 Qwen3 + model-level offloading is optimized for 8GB VRAM.

📚 References

• Original Model: Tongyi-MAI/Z-Image-Turbo
• LightX2V: GitHub
• Qwen3-4B-GPTQ-Int4: JunHowie/Qwen3-4B-GPTQ-Int4

🤝 Community

If you find this project helpful, please give us a ⭐ on GitHub