LightX2V

A lightweight inference framework for image and video generation, with distilled models, quantization and many hardware back ends.

  • Serve at scale
  • Apache 2.0 · Open source
  • Linux
  • Verified

LightX2V is an inference framework from the ModelTC team for text-to-video, image-to-video, text-to-image and image-editing models. The README describes it as lightweight and aimed at efficient, high-performance synthesis. It supports quantized weights, step-distilled models and lightweight autoencoders, and it adds parallelism and CPU offload so a model can be split across GPUs or squeezed onto a smaller one. The recommended way to use it is a server with a request API, which listens on port 8000 by default, started from a Docker environment.

Its model list covers video models such as Wan 2.1 and 2.2, LTX-2.5, MiniMax H3 and HunyuanVideo-1.5, and image models such as Qwen-Image 2.1, FLUX.2, Z-Image, HiDream-O1-Image, ERNIE-Image and MiniMax-H3-Image, along with restoration models like SeedVR2. Its newest news item, dated 11 November 2026, adds MiniMax-H3-Image with tensor and sequence parallelism and CPU offload. The project also publishes distilled LoRAs, and for its distilled Wan 2.2 14B build on one RTX 5090 it reports more than 50 times the end-to-end speed of the original, a figure that comes from its own model-card test conditions.

The code is Apache 2.0 licensed. Beyond NVIDIA GPUs, the news mentions deployments on AMD ROCm, Ascend 910B, Moore Threads MUSA, Hygon DCU, Cambricon, Intel AI PCs and other accelerators. The same team also released a separate LightX2V APP for running selected image models on a MacBook. Treat it as a deployment tool for people who already know which model they want to serve.

Good for

  • Serving Wan, LTX, MiniMax H3 and Qwen-Image models with step-distilled weights for faster generation
  • Running video generation on a card with less memory through offload and quantization
  • Deployments on non-NVIDIA accelerators that the README lists

Watch out for

  • Each model has its own guide, weights and server script, so setup takes careful reading
  • Speedups quoted in the README are the maintainers' own measurements under their test conditions
  • Mac support comes from a separate LightX2V APP that runs only selected image models

Links