Hy1.5-Quantized-Models

🎬 Hy1.5-Quantized-Models

<img src="https://raw.githubusercontent.com/ModelTC/LightX2V/main/assets/img_lightx2v.png" width="75%" />

---

🤗 HuggingFace | GitHub | License

---

This repository contains quantized models for HunyuanVideo-1.5 optimized for use with LightX2V. These quantized models significantly reduce memory usage while maintaining high-quality video generation performance.

📋 Model List

DIT (Diffusion Transformer) Models

• hy15_720p_i2v_fp8_e4m3_lightx2v.safetensors - 720p Image-to-Video quantized DIT model
• hy15_720p_t2v_fp8_e4m3_lightx2v.safetensors - 720p Text-to-Video quantized DIT model

Encoder Models

• hy15_qwen25vl_llm_encoder_fp8_e4m3_lightx2v.safetensors - Quantized text encoder (Qwen2.5-VL LLM Encoder)

🚀 Quick Start

Installation

First, install LightX2V:

pip install -v git+https://github.com/ModelTC/LightX2V.git

Or build from source:

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

Download Models

Download the quantized models from this repository:

# Using git-lfs
git lfs install
git clone https://huggingface.co/lightx2v/Hy1.5-Quantized-Models

# Or download individual files using huggingface-hub
pip install huggingface-hub
python -c "from huggingface_hub import hf_hub_download; hf_hub_download(repo_id='lightx2v/Hy1.5-Quantized-Models', filename='hy15_720p_i2v_fp8_e4m3_lightx2v.safetensors', local_dir='./models')"

💻 Usage in LightX2V

Text-to-Video (T2V) Example

from lightx2v import LightX2VPipeline

# Initialize pipeline
pipe = LightX2VPipeline(
    model_path="/path/to/hunyuanvideo-1.5/",  # Original model path
    model_cls="hunyuan_video_1.5",
    transformer_model_name="720p_t2v",
    task="t2v",
)

# Enable quantization
pipe.enable_quantize(
    quant_scheme='fp8-sgl',
    dit_quantized=True,
    dit_quantized_ckpt="/path/to/hy15_720p_t2v_fp8_e4m3_lightx2v.safetensors",
    text_encoder_quantized=True,
    text_encoder_quantized_ckpt="/path/to/hy15_qwen25vl_llm_encoder_fp8_e4m3_lightx2v.safetensors",
    image_encoder_quantized=False,
)

# Optional: Enable offloading for lower VRAM usage
pipe.enable_offload(
    cpu_offload=True,
    offload_granularity="block",  # For HunyuanVideo-1.5, only "block" is supported
    text_encoder_offload=True,
    image_encoder_offload=False,
    vae_offload=False,
)

# Optional: Use lighttae
pipe.enable_lightvae(
    use_tae=True,
    tae_path="/path/to/lighttaehy1_5.safetensors",
    use_lightvae=False,
    vae_path=None,
)

# Create generator
pipe.create_generator(
    attn_mode="sage_attn2",
    infer_steps=50,
    num_frames=121,
    guidance_scale=6.0,
    sample_shift=9.0,
    aspect_ratio="16:9",
    fps=24,
)

# Generate video
seed = 123
prompt = "A beautiful sunset over the ocean with waves gently crashing on the shore."
negative_prompt = ""
save_result_path="/path/to/output.mp4"

pipe.generate(
    seed=seed,
    prompt=prompt,
    negative_prompt=negative_prompt,
    save_result_path=save_result_path,
)

Image-to-Video (I2V) Example

from lightx2v import LightX2VPipeline

# Initialize pipeline
pipe = LightX2VPipeline(
    model_path="/path/to/hunyuanvideo-1.5/",  # Original model path
    model_cls="hunyuan_video_1.5",
    transformer_model_name="720p_i2v",
    task="i2v",
)

# Enable quantization
pipe.enable_quantize(
    quant_scheme='fp8-sgl',
    dit_quantized=True,
    dit_quantized_ckpt="/path/to/hy15_720p_i2v_fp8_e4m3_lightx2v.safetensors",
    text_encoder_quantized=True,
    text_encoder_quantized_ckpt="/path/to/hy15_qwen25vl_llm_encoder_fp8_e4m3_lightx2v.safetensors",
    image_encoder_quantized=False,
)

# Optional: Use lighttae
pipe.enable_lightvae(
    use_tae=True,
    tae_path="/path/to/lighttaehy1_5.safetensors",
    use_lightvae=False,
    vae_path=None,
)

# Optional: Enable offloading for lower VRAM usage
pipe.enable_offload(
    cpu_offload=True,
    offload_granularity="block",
    text_encoder_offload=True,
    image_encoder_offload=False,
    vae_offload=False,
)

# Create generator
pipe.create_generator(
    attn_mode="sage_attn2",
    infer_steps=50,
    num_frames=121,
    guidance_scale=6.0,
    sample_shift=7.0,
    fps=24,
)

# Generate video
seed = 42
prompt = "The image comes to life with smooth motion and natural transitions."
negative_prompt = ""
save_result_path="/path/to/output.mp4"

pipe.generate(
    seed=seed,
    image_path="/path/to/input_image.jpg",
    prompt=prompt,
    negative_prompt=negative_prompt,
    save_result_path=save_result_path,
)

⚙️ Quantization Scheme

These models use FP8-E4M3 quantization with the SGL (SGLang) kernel scheme (fp8-sgl). This quantization format provides:

• Significant memory reduction: Up to 50% reduction in VRAM usage
• Maintained quality: Minimal quality degradation compared to full precision models
• Faster inference: Optimized kernels for accelerated computation

Requirements

To use these quantized models, you need to install the SGL kernel:

# Requires torch == 2.8.0
pip install sgl-kernel --upgrade

Alternatively, you can use VLLM kernels:

pip install vllm

For more details on quantization schemes, please refer to the LightX2V Quantization Documentation.

📊 Performance Benefits

Using quantized models provides:

• Lower VRAM Requirements: Enables running on GPUs with less memory (e.g., RTX 4090 24GB)
• Faster Inference: Optimized quantized kernels accelerate computation
• Quality Preservation: FP8 quantization maintains high visual quality

🔗 Related Resources

• LightX2V GitHub Repository
• LightX2V Documentation
• HunyuanVideo-1.5 Original Model
• LightX2V Examples

📝 Notes

• Important: All advanced configurations (including enable_quantize()) must be called before create_generator(), otherwise they will not take effect.
• The original HunyuanVideo-1.5 model weights are still required. These quantized models are used in conjunction with the original model structure.
• For best performance, we recommend using SageAttention 2 (sage_attn2) as the attention mode.

🤝 Citation

If you use these quantized models in your research, please cite:

@misc{lightx2v,
  author = {LightX2V Contributors},
  title = {LightX2V: Light Video Generation Inference Framework},
  year = {2025},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/ModelTC/lightx2v}},
}

📄 License

This model is released under the Apache 2.0 License, same as the original HunyuanVideo-1.5 model.