🎬 LightLingBot-Video

Few-step distilled LoRA checkpoints for LingBot-Video, accelerated by LightX2V.

LightLingBot-Video provides distilled LoRA checkpoints for fast LingBot-Video inference. The same LoRA configuration supports text-to-video (T2V), text-to-image (T2I), and image-to-video (I2V) generation.

📋 Table of Contents

• Features
• Installation
• Model Preparation
• Inference
• Generation Results
• Notes
• Acknowledgements

✨ Features

• ⚡ 4-step inference with distilled LoRA checkpoints
• 🎬 T2V, T2I, and I2V generation
• 🔧 One unified LoRA configuration shared by all tasks
• 🚀 LightX2V integration for efficient inference and deployment
• 🎯 Two distilled LoRA versions for comparison

📦 Installation

git clone https://github.com/ModelTC/LightX2V.git
cd LightX2V
python3 -m pip install -e .

Follow the LightX2V installation guide to install the optional acceleration dependencies required by your hardware.

📥 Model Preparation

Download the LingBot-Video base model and the distilled LoRA repository:

huggingface-cli download robbyant/lingbot-video-moe-30b-a3b \
  --local-dir /path/to/LingBot-Video

huggingface-cli download lightx2v/LightLingBot-Video \
  --local-dir /path/to/LightLingBot-Video

Before inference:

1. Set model_path in the corresponding script under scripts/lingbot_video/.
2. Set the LoRA checkpoint path in configs/lingbot_video/lingbot_video_lora.json.
3. Adjust CUDA_VISIBLE_DEVICES in the script if needed.
4. For I2V, also set image_path in the I2V script.

🚀 Inference

Run the following commands from the LightX2V repository root. These commands use the shell scripts under scripts/lingbot_video; no examples entry point is required.

Text-to-Video

bash scripts/lingbot_video/run_lingbot_video_t2v.sh

Text-to-Image

bash scripts/lingbot_video/run_lingbot_video_t2i.sh

Image-to-Video

bash scripts/lingbot_video/run_lingbot_video_i2v.sh

Recommended Settings

Setting	Base model	Distilled LoRA
Inference steps	40	4
LoRA strength	—	1.0
Resolution	832 × 480	832 × 480
Guidance scale	3.0	1.0
Sample shift	3.0	3.0

🎬 Generation Results

Text-to-Video (T2V)

Method	Sample 1	Sample 2	Sample 3
40 steps
4-step LoRA v1
4-step LoRA v2

Text-to-Image (T2I)

Method	Sample 1	Sample 2	Sample 3	Sample 4
40 steps
4-step LoRA v1
4-step LoRA v2

Image-to-Video (I2V)

Method	Sample 1	Sample 2	Sample 3	Sample 4
40 steps
4-step LoRA v1
4-step LoRA v2

⚠️ Notes

• The distilled LoRA configuration uses 4 inference steps and disables CFG.
• Merged LoRA inference requires original, non-quantized DiT weights.
• Prompts and negative prompts should follow the structured JSON format used in scripts/lingbot_video/.
• T2V, T2I, and I2V share configs/lingbot_video/lingbot_video_lora.json.
• Speed and memory usage depend on the GPU, attention backend, resolution, and video length.

🤝 Acknowledgements

• LingBot-Video
• LightX2V

If you find this project helpful, please give LightX2V a ⭐.