LightX2V

[ English | 中文 ]

---

LightX2V is an advanced lightweight image/video generation inference framework engineered to deliver efficient, high-performance image/video synthesis solutions. This unified platform integrates multiple state-of-the-art image/video generation techniques, supporting diverse generation tasks including text-to-video (T2V), image-to-video (I2V), text-to-image (T2I), image-editing (I2I). X2V represents the transformation of different input modalities (X, such as text or images) into vision output (Vision).

Quick Start · Models · Documentation & Deployment

🌐 Try the LightX2V-powered AI avatar and multimodal content production platform online: LightX2V Studio

🍎 MacBook local inference app: LightX2V APP

🤗 HuggingFace Model Repository: LightX2V HuggingFace

🌟 Developer Quick Start: LightX2V Developer Quick Start Guide

📝 Technical Blog: LightX2V Blog

👋 Join our WeChat group! LightX2V Robot WeChat ID: Jo-1212liz

🧾 Community Code Contribution Guidelines

Before submitting, please ensure that the code format conforms to the project standard. You can use the following execution command to ensure the consistency of project code format.

pip install ruff pre-commit
pre-commit run --all-files

We thank the following community developers for their contributions to LightX2V:

zhtshr · triple-Mu · vivienfanghuagood · yeahdongcn · kikidouloveme79 · ziyanxzy · Tyr0727 · hufangjian2017 · Fatemanx · qiuxin2012

:fire: Latest News

• October 26, 2026: 🚀 Added autoregressive inference for TaoMate-H3, StreamRender-H3, Matrix-Game-4, LingBot-Fast, and Realtime-Video. Model-specific scripts and BF16/FP8 profiles are available in the linked directories.

• October 18, 2026: 🚀 We now support EdenLive for video generation controlled by text, a first-frame image, and a proxy recording. Wan2.1 14B and Wan2.2 5B each support bidirectional and four-step autoregressive inference, with realtime profiles for RTX 5090 and H100/H200. See the deployment guide for model conversion, quantization, and streaming inference.

• October 13, 2026: 🎉 We released the LightX2V APP, which supports local inference on MacBook with Qwen-Image-2.1 Turbo, PE-T2I GGUF Q8_0, and PE-I2I GGUF Q8_0 models.

• September 22, 2026: 🚀 We added and optimized support for SwiftVR, covering image and video restoration and super-resolution. See the usage guide.

• September 20, 2026: 🚀 We added day-0 support for Qwen-Image-2.1, covering text-to-image generation and prompt-based image editing. See the usage guide.

• August 27, 2026: 🚀 We released the MiniMax-H3 Turbo 8-step 768p distilled LoRA for generating 768p video with synchronized audio in 8 inference steps using MiniMax-H3. See the MiniMax-H3 guide.

• August 11, 2026: 🚀 We released the MiniMax-H3 Turbo 4-step 768p distilled LoRA and added support for generating 768p video with synchronized audio in 4 inference steps without CFG. See the MiniMax-H3 guide.

• August 7, 2026: 🚀 We added support for synchronized video and audio generation with MiniMax-H3 and released a Prompt Rewriter LoRA that expands short prompts into structured descriptions of visuals, soundscapes, and music. See the MiniMax-H3 guide.

• July 28, 2026: 🚀 We release the LightLingBot-Video 4-step distilled LoRA for LingBot-Video. It supports T2V, T2I, and I2V generation in only 4 inference steps without CFG. See the LingBot-Video inference scripts for usage.

• July 23, 2026: ⚡️ We release the blog: LightX2V ROS: Closing the Loop for Action-Generating World Models

• July 19, 2026: ⚡️ We release the blog: Wan2.2-NVFP4-Sparse: Extremely Fast Wan 2.2 14B Inference

• June 15, 2026: 🚀 Supported deployment on T-head PPU.

• May 29, 2026: 🚀 We introduce an extremely efficient Wan 2.2 14B variant (T2V and I2V): NVFP4 Quantization-Aware Step Distillation with Sparse Attention for Blackwell Architecture. Under the test conditions in the model card, inference with 4 NFE achieves over 50× end-to-end speedup over the original model with 40 NFE on a single RTX 5090 GPU.

• April 30, 2026: 🚀 We now support deployment on iluvatar. Thanks to the iluvatar team.

• April 20, 2026: 🚀 We are excited to release the Wan2.2-I2V-A14B-4step-720p-high and Wan2.2-I2V-A14B-4step-720p-low models. Compared to previous iterations, this version was trained on a high-quality 720p dataset and features an optimized low-noise training algorithm. These enhancements significantly boost the model's performance in fine-grained detail rendering and visual texture.

• April 17, 2026: 🚀 We support the WorldMirror 2.0 model. On a single H100 GPU, LightX2V delivers approximately 1.2× speedup. For detailed usage instructions, please refer to this guide.

• April 10, 2026: 🎉 We released a technical blog on disaggregated deployment

  • LightX2V Disaggregated Deployment: Breaking the Memory and Throughput Bottlenecks of Diffusion Model Inference

• March 5, 2026: 🚀 We now support deployment on Intel AIPC PTL. Thanks to the Intel team.

• March 5, 2026: 🚀 We now support disaggregated deployment based on Mooncake. More improvements and documentation for disaggregated deployment are in progress. Thanks to the Mooncake team for their help!

• February 27, 2026: 🚀 We now support FP8 and NVFP4 quantization for autoregressive video generation models (Self Forcing)! You can find the quantized model here: Self-Forcing-FP8, Self-Forcing-NVFP4.

• February 11, 2026: 🎉 We are excited to announce GenRL - a scalable reinforcement learning framework for visual generation! GenRL enables training diffusion/flow models with multi-reward optimization (HPSv3, VideoAlign, etc.) using GRPO algorithm. We've released high-performance LoRA checkpoints trained with multi-node multi-GPU setup, demonstrating significant improvements in aesthetic quality, motion coherence, and text-video alignment. Check out our model collection on HuggingFace! Give us a ⭐ if you find it useful!

• January 20, 2026: 🚀 We support the LTX-2 audio-video generation model, featuring CFG parallelism, block-level offload, and FP8 per-tensor quantization. Usage examples can be found in scripts/ltx2/inference and scripts/ltx2.

• January 6, 2026: 🚀 We updated the 8-step CFG/step-distilled models for Qwen-Image-2512 and Qwen/Qwen-Image-Edit-2511. You can download the corresponding weights from Qwen-Image-Edit-2511-Lightning and Qwen-Image-2512-Lightning for use. Usage tutorials can be found here.

• January 6, 2026: 🚀 Supported deployment on Enflame S60 (GCU).

• December 31, 2025: 🚀 We support the Qwen-Image-2512 text-to-image model since Day 0. Our HuggingFace has been updated with CFG / step-distilled LoRA. Usage examples can be found here.

• December 27, 2025: 🚀 Supported deployment on MThreads MUSA.

• December 25, 2025: 🚀 Supported deployment on AMD ROCm and Ascend 910B.

• December 23, 2025: 🚀 We support the Qwen-Image-Edit-2511 image editing model since Day 0. On a single H100 GPU, LightX2V delivers approximately 1.4× speedup. We support for CFG parallelism, Ulysses parallelism, and efficient offloading technologies. Our HuggingFace has been updated with CFG / step-distilled LoRA and FP8 weights. Usage examples can be found here. Combined with LightX2V, 4-step CFG / step distillation, and the FP8 model, the maximum acceleration can reach up to approximately 42×. Feel free to try LightX2V Online Service with Image to Image and Qwen-Image-Edit-2511 model.

• December 22, 2025: 🚀 Added Wan2.1 NVFP4 quantization-aware 4-step distilled models; weights are available on HuggingFace: Wan-NVFP4.

• December 15, 2025: 🚀 Supported deployment on Hygon DCU.

• December 4, 2025: 🚀 Supported GGUF format model inference & deployment on Cambricon MLU590/MetaX C500.

• November 24, 2025: 🚀 We released 4-step distilled models for HunyuanVideo-1.5! These models enable ultra-fast 4-step inference without CFG requirements, achieving approximately 25x speedup compared to standard 50-step inference. Both base and FP8 quantized versions are now available: Hy1.5-Distill-Models.

• November 21, 2025: 🚀 We support the HunyuanVideo-1.5 video generation model since Day 0. With the same number of GPUs, LightX2V can achieve a speed improvement of over 2 times and supports deployment on GPUs with lower memory (such as the 24GB RTX 4090). It also supports CFG/Ulysses parallelism, efficient offloading, TeaCache/MagCache technologies, and more. We will soon update more models on our HuggingFace page, including step distillation, VAE distillation, and other related models. Quantized models and lightweight VAE models are now available: Hy1.5-Quantized-Models for quantized inference, and LightTAE for HunyuanVideo-1.5 for fast VAE decoding. Refer to this for usage tutorials, or check out the Python API guide for code examples.

💡 Quick Start

We recommend using LightX2V through its service API. The examples below cover Qwen-Image-2.1 text-to-image (T2I) generation and MiniMax-H3 text-to-audio-video (T2AV) generation.

Environment Setup

We recommend the Docker environment. Clone the repository in your prepared environment:

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

Run these commands in a prepared Python environment. The service example uses scripts from the cloned repository.

Installation from Git

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

Installation from Source (Repository Root)

uv pip install -v . # or: pip install -v .

See the environment setup guide for attention and quantization kernel installation.

Usage Examples

Both servers use port 8000 by default. Stop the running server before switching examples.

Qwen-Image-2.1: Text-to-Image

Follow the Qwen-Image-2.1 guide to prepare the weights, then set the repository path, model path, and GPU in the server script. Run the commands below from the repository root.

# Terminal 1: start the server
bash scripts/qwen_image_21/server/start_server.sh

After the model loads, submit the example request from another terminal:

# Terminal 2: submit a T2I request
python scripts/qwen_image_21/server/post_t2i.py

Edit the prompt, size, and seed in the request script to generate your own images. Once the request completes, the image is saved by default to save_results/qwen_image_21_t2i.png under the repository root.

MiniMax-H3: Text-to-Audio-Video

Follow the MiniMax-H3 guide to prepare the weights and initial AdaLN cache, and set the repository path, model path, and GPU in the server script. Run the commands below from the repository root.

# Terminal 1: start the server
bash scripts/minimax_h3/server/start_server.sh

After the model loads, submit the example request from another terminal:

# Terminal 2: submit a T2AV request
python scripts/minimax_h3/server/post_t2av.py

Edit the prompt in the request script to generate your own content. Requests are asynchronous; results are saved by default to save_results/server_cache/outputs/ under the repository root.

🤖 Supported Model Ecosystem

LightX2V supports image, video, and world models and provides quantized weights, distilled models, and lightweight autoencoders.

Models / Variants and Weights	Guides / Scripts
MiniMax-H3 (causal RefA2V requires a separate checkpoint)	Guide · Causal
TaoMate-H3 (autoregressive T2AV)	Run script
StreamRender-H3 (autoregressive RefV2V)	Run script
LTX-2 / LTX-2.3 / LTX-2.5 (distilled)	Guide
HunyuanVideo-1.5	Guide
Wan2.1 / Wan2.2	Wan2.1 · Wan2.2
Wan2.1-VACE / Wan2.2-VACE-Fun-A14B	Wan2.1 · Wan2.2
Wan2.2-S2V	Guide
Wan2.2-Animate / Wan2.2-Animate-2 (distilled)	Guide
LingBot-Video Dense 1.3B / MoE 30B-A3B	Scripts
SekoTalk (including the autoregressive variant)	Scripts
InfiniteTalk	Scripts
Wan-Dancer	Scripts
Self-Forcing	Guide

Models / Variants and Weights	Guides / Scripts
Qwen-Image / Qwen-Image-2512	Guide
Qwen-Image-Edit / 2509 / 2511	Guide
Qwen-Image-2.1 / Turbo	Guide
Qwen-Image-Layered	Guide
FLUX.2 dev / klein (base and distilled klein variants)	Scripts
HunyuanImage-3.0-Instruct	Scripts
Z-Image / Z-Image-Turbo	Turbo guide; base model: enable CFG (guidance > 1) and adjust sampling steps
LongCat-Image / Edit / Edit-Turbo	Guide; Edit-Turbo requires its own weights, 8 steps, and CFG disabled
BAGEL	Guide
HiDream-O1-Image / Dev-2604	Scripts
ERNIE-Image / Turbo	Scripts
SenseNova-U1 (NeoPP backend, used with LightLLM)	Guide

Models / Variants and Weights	Guides / Scripts
SeedVR2	Guide
SwiftVR	Guide
RelayVSR	Guide

Models / Variants and Weights	Guides / Scripts
Cosmos3-Nano / Super (including Nano-Policy-DROID)	Scripts
LingBot-World / Fast	Scripts
Matrix-Game-4 (Wan2.2 5B Control; autoregressive video rerendering)	Run script
LingBot-Fast (autoregressive I2V)	Run script
Realtime-Video (Wan2.1 autoregressive T2V)	Run script
WorldPlay (bidirectional, autoregressive, and distilled)	Guide
EdenLive (Wan2.1 14B / Wan2.2 5B; bidirectional and four-step autoregressive)	Guide
WorldMirror 2.0 (3D reconstruction)	Guide
Depth Anything 3 MONO-LARGE (image/video relative depth)	Guide and run script
DreamZero	Scripts
LingBot-VA	Scripts
Motus	Scripts
Fast-WAM	Scripts
RealtimeWAM (based on Fast-WAM / Faster-WAM)	Guide
Hunyuan3D-2.1 (image-to-3D shape generation)	Scripts

Models / Variants and Weights	Guides / Scripts
MiniMax-H3 Turbo — 4-/8-step distilled LoRAs, including 768p T2AV / FL2AV and 4-step Ref2AV	Guide
LightLingBot-Video — 4-step distilled LoRA for T2V, T2I, and I2V	Scripts
Wan2.1-Distill-Models / Wan2.2-Distill-Models	Wan2.1 · Wan2.2
Wan2.1-Distill-Loras / Wan2.2-Distill-Loras	Wan2.1 · Wan2.2
Wan-NVFP4 / LightWan2.2-A14B	Wan2.1 · Wan2.2
Hy1.5-Distill-Models / Hy1.5-Quantized-Models	Guide
Qwen-Image-Lightning / Qwen-Image-2512-Lightning / Qwen-Image-2.1-Lightning	Qwen-Image · 2.1
Qwen-Image-Edit-2511-Lightning	Guide
Z-Image-Turbo-Quantized	Guide
Self-Forcing-FP8 / Self-Forcing-NVFP4	Guide
Autoencoders (LightVAE / LightTAE)	Python API

🚀 Core Features

• Few-Step Inference: Support distilled checkpoints and distillation LoRAs, including DMD2 and Wan2.1 MeanFlow inference.
• Quantization & Optimized Kernels: Support W8A8 (INT8 / FP8) and W4A4 (NVFP4) quantization, with optimized attention backends such as FlashAttention and SageAttention.
• Hierarchical Offloading: Manage weights across disk, CPU, and GPU with model-level and block-level offloading.
• Feature Caching: Support TeaCache, MagCache, and TaylorSeer to reuse intermediate results across denoising steps.
• Parallel Inference: Support tensor parallelism (TP), sequence parallelism (SP), CFG parallelism, and hybrid configurations.
• Variable-Resolution Inference: Schedule latent resolutions across denoising stages to control computation at each stage.

Available optimizations depend on the model, hardware, and selected configuration.

📚 Documentation & Deployment

Use Case	Documentation & Examples
Environment Setup & Configuration	Quick Start Guide
Model Preparation & Loading	MiniMax-H3 · Wan2.1 / VACE · Model Formats & Loading
Service Deployment & API Usage	Service Deployment · API Reference
Interactive Interfaces & Workflow Integration	Gradio-Lightx2v · ComfyUI
Local Deployment on Windows	Windows Deployment Guide
Resource-Constrained and Low-Latency Deployment	Low-Resource Deployment · Low-Latency Deployment
Model Quantization & LoRA Deployment	Quantization · LoRA Deployment
Inference Performance Optimization	Weight Offloading · Feature Caching · Attention Kernels · Parallel Inference · Variable-Resolution Inference · Step Distillation
Wan2.1 NVFP4 4-Step Inference	I2V Example · T2V Example · Quantization Kernel Installation
Code Examples & Development References	Python API · scripts

🤝 Acknowledgments

We thank the following open-source projects and communities for the models, tools, and research that have helped shape LightX2V:

Tencent-Hunyuan · Wan-Video · Qwen-Image · MiniMax-H3 · FlashAttention · SageAttention · LightLLM · sglang · vllm · flashinfer · MagiAttention · radial-attention · xDiT · FastVideo · Mooncake

We also thank the cloud inference platforms supporting the wider ecosystem:

• Atlas Cloud — a full-modal AI inference platform whose hosted API serves the Wan / Seedance / Kling model families, for teams that prefer a managed API over self-hosting.

• Sensecore — Sensecore builds efficient, low-cost, scalable AI cloud infrastructure and develops deep learning platforms, algorithms, and models to support AI innovation in industry and academia.

✏️ Citation

If you find LightX2V useful in your research, please consider citing our work:

@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}},
}

📞 Contact & Support

For questions, suggestions, or support, please feel free to reach out through:

• 🐛 GitHub Issues - Bug reports and feature requests

---