Note: What Plugs Where page contains some information on how native and wrapper workflows for VACE are constructed. This page contains additional/advanced examples.
drozbay has shared the following advanced VACE workflows relying on his ComfyUI-WanVaceAdvanced node pack, with ClownShark and VACE:
DWPose Estimator and Depth Anything V2 as 2 separate VACE control videos
Canny Edge limited by depth masking: “keep the outlines, … don’t … details inside the objects”Note: WanVacePhantomDualV2 from WanVaceAdvanced is compatible with Phantom but not linked to it otherwise.
vace_reference and phantom_images on this node are fundamentally different: “they go to different model layers and they are treated differently”.
vace is still pretty iffy when trying to use the reference input for style it’s not a style-transfer model, even if it occasionally does work the reference is really for either background, setting, or subject(s)
it will not be great for character unless your character reference is pretty much exactly where the character will be in the first frame. In that case I recommend you use Phantom
and don’t use the skeleton+depth mix; send in the depth and the skeleton into separate Vace control inputs; either chain WanVaceToVideo nodes, or use WanVacePhantomDualV2 node
Q: load FUN VACE module for High Model first, and then VACE 2.1 for the LOW model? A: sometimes
can try WAN 2.1 + Wan 2.1 to Wan 2.2 LOW LoRA + VACE 2.1 on LOW
original Vace was pretty weak with pose control