I've put together a simple custom 2 part text-to-video and image-to-video workflow with VACE, that uses the self_forcing_dmd model and Kijai's WAN 14B Lightx2v LoRA which brings down inference to only 4 steps. To make it more accessible for those of us with less powerful GPU's, it also uses a lightweight UMT5 GGUF clip model.

If you would like a full breakdown of the workflow and the installation process, you can watch my YouTube tutorial below.

Link -  https://youtu.be/TyiEHj8TtTE