--- license: apache-2.0 library_name: lightx2v base_model: - robbyant/lingbot-video-moe-30b-a3b pipeline_tag: text-to-video tags: - video-generation - text-to-video - image-to-video - text-to-image - distillation - lora --- # 🎬 LightLingBot-Video > Few-step distilled LoRA checkpoints for LingBot-Video, accelerated by > [LightX2V](https://github.com/ModelTC/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](#-features) - [Installation](#-installation) - [Model Preparation](#-model-preparation) - [Inference](#-inference) - [Generation Results](#-generation-results) - [Notes](#-notes) - [Acknowledgements](#-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 ```bash git clone https://github.com/ModelTC/LightX2V.git cd LightX2V python3 -m pip install -e . ``` Follow the [LightX2V installation guide](https://github.com/ModelTC/LightX2V) to install the optional acceleration dependencies required by your hardware. ## 📥 Model Preparation Download the LingBot-Video base model and the distilled LoRA repository: ```bash 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 bash scripts/lingbot_video/run_lingbot_video_t2v.sh ``` ### Text-to-Image ```bash bash scripts/lingbot_video/run_lingbot_video_t2i.sh ``` ### Image-to-Video ```bash 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 |

| | 4-step LoRA v1 | T2I LoRA v1 sample 1

| | 4-step LoRA v2 | T2I LoRA v2 sample 1

| ### 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](https://huggingface.co/inclusionAI/LingBot-Video) - [LightX2V](https://github.com/ModelTC/LightX2V) If you find this project helpful, please give [LightX2V](https://github.com/ModelTC/LightX2V) a ⭐.