alx-code/lingbot-video-1.3b-fp8

This project provides a ComfyUI add on for running LingBot-Video Dense 1.3B, an AI model that turns text or a starting image into a short video. ComfyUI is a visual tool people use to build and run AI image and video pipelines by connecting nodes together, and this project supplies the custom nodes needed to load and run this particular video model inside it, along with three ready made workflows: generating video from a text description, generating video starting from one image, and an experimental version that also uses a final image. The main technical contribution here is a modified version of the model that runs some of its internal calculations using a lower precision number format called FP8 instead of the usual higher precision format, which can make certain steps run faster on supported graphics cards. The project only applies this lower precision to specific parts of the model where it measured a real speed benefit, keeping everything else at the original, higher quality precision to avoid noticeably hurting output quality. In the author's own test on a single RTX 5080 graphics card, this made one sampling step about 27 percent faster than the standard version, though the author is careful to note this was one narrow smoke test, not a broad benchmark across different hardware. Setting this up is involved: it requires a specific graphics card capable of running this lower precision format, a particular combination of Python, PyTorch, and CUDA versions, and downloading both the original multi gigabyte base model and this project's separate lower precision component. A handful of other optional add ons can be installed for extra features like upscaling the video or smoothing motion between frames. Detailed installation commands, checksums to verify the downloads, and known limitations are documented in the project's files. This is described as an independent community project, not an official release from the company that made the underlying model.