WanVideo Extra Model Select

Some capabilities in the Wan ecosystem don't come baked into the base checkpoint - VACE (reference-to-video, extension, inpainting) is the obvious one, and it's normally its own separate model that gets combined with a base Wan model rather than swapped in for it. This node is how WanVideoWrapper handles that combination: pick an extra model, and it gets added to whatever main model you're already running, rather than replacing it.

What it's for

VACE is Alibaba's own official control framework for Wan - reference images plus a pose or depth control video, driving motion-consistent generation - and it's the ecosystem's dominant control mechanism by a wide margin, ahead of ControlNet ports or LoRA-based control. The other option this node's description calls out, MTV Crafter, is a motion-transfer model rather than a control framework proper. Either way, the pattern is the same: it's an extra model, loaded and added on top of the diffusion model you already have selected elsewhere in your graph, rather than something you load standalone.

Because the optional prev_model input takes the same VACEPATH type this node outputs, you can chain multiple WanVideo Extra Model Select nodes together - select VACE, feed it as prev_model into a second instance selecting something else, and so on - building up a combined stack of extras rather than being limited to one.

The inputs and output that matter

extra_model - the dropdown of available extra models, populated from what you've downloaded into ComfyUI/models/diffusion_models. prev_model (VACEPATH, optional, defaults to none) - chain in a previously selected extra model here if you're combining more than one.

Output is a single extra_model (VACEPATH), which feeds into whatever node in your graph combines it with your main Wan model - typically your model loader's own extra-model input, if it has one, or a dedicated combine node elsewhere in the pack.

How to install it

Via ComfyUI Manager, search WanVideoWrapper and install, or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper

then restart. The extra models themselves - VACE weights, or whatever else you want to stack - are separate downloads that go into ComfyUI/models/diffusion_models, same folder as your base Wan checkpoints. VACE ships in both 1.3B and 14B sizes; match the size to your base model and your VRAM budget, since the KB's numbers put 14B VACE at 16-24GB+ depending on how much you're willing to trade in speed for headroom.

Common issues & troubleshooting

Empty dropdown, nothing to select. This is a loader, not a downloader - if extra_model shows no options, you haven't put any compatible checkpoint into ComfyUI/models/diffusion_models yet.

Combining an extra model with a base model of a mismatched scale. VACE 14B paired with a 1.3B base model (or vice versa) is a mismatch the KB's own VRAM notes warn about generally - dimensions and expected quality don't line up cleanly across scales in this ecosystem, so keep your base model and your extra model in the same weight class unless you specifically know the combination is supported.

VACE generation is slow even after this node loads cleanly. That's expected and not this node's fault - VACE is genuinely one of the slower paths in the Wan ecosystem at full precision. The community's standard fix is a distillation LoRA like CausVid on top, which drops VACE to 2-4 steps at CFG 1.0 rather than trying to speed up the extra-model loading itself.