Dummy Comfy Wan Model Object

The name sounds like a joke, but it's doing something genuinely useful: it hands ComfyUI's native BasicScheduler node a fake Wan model just so BasicScheduler will compute the right sigma schedule. That's it. It's plumbing, not a generator - you'll never see this node in a finished video, only in the graph that builds toward one.

Why this exists

WanVideoWrapper mostly lives in its own world - its own model loader, its own sampler, its own conditioning types - because that's how kijai gets the finer control (quantization, offloading, block-level LoRA editing) that native ComfyUI's Wan support doesn't expose. But ComfyUI's built-in BasicScheduler node is genuinely good at one thing: turning a shift value and a step count into a correct sigma sequence for flow-matching sampling. Wan is a flow-matching model, and getting the sigma schedule right matters for how the noise steps down over your sampling run.

The catch is BasicScheduler expects a real MODEL object as input, because in native ComfyUI the model carries the sampling config. If you're building a Wan graph that stays inside WanVideoWrapper's own nodes, you don't have one of those lying around - you have WanVideoWrapper's own model type instead. DummyComfyWanModelObject bridges the gap: it fabricates a minimal, empty model object that's just enough to satisfy BasicScheduler and get sigmas out, without the cost or hassle of loading an actual multi-gigabyte checkpoint for a value you could compute from a formula.

The input and output that matter

There's exactly one input: shift (default 1, range −100 to 100). This is the same flow-matching shift parameter you'll see across WanVideoWrapper's samplers - a higher shift concentrates more of the schedule's resolution toward the high-noise end, which changes how aggressively structure gets laid down early versus refined late. If you're feeding this into BasicScheduler to reproduce a specific sampler's behavior, match the shift value to what that sampler actually uses (WanVideoDiffusionForcingSampler, for instance, defaults to 8).

The output is a single model - the dummy object - which you wire straight into BasicScheduler's model input.

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. This node itself downloads nothing and needs no model files - it's the one node in the pack that's specifically designed to avoid loading a model.

Common issues & troubleshooting

You don't need this node for a normal WanVideoWrapper workflow. If you're using the pack's own samplers end to end, they compute their own sigma schedule internally from the shift and scheduler settings on the sampler node itself. Reach for DummyComfyWanModelObject specifically when you're mixing graphs - pulling BasicScheduler or another native-ComfyUI scheduling node into an otherwise WanVideoWrapper pipeline - and need a model-shaped object to unblock it.

Mismatched shift values produce a schedule that doesn't match your actual sampler. Since this node's whole purpose is reproducing a specific schedule, double-check the shift here against whatever sampler you're trying to match. Get it wrong and BasicScheduler will happily hand you sigmas for a different sampling behavior than the one you're running - no error, just a schedule that's quietly off.

It's not a substitute for loading your real Wan model. The object it produces is empty by design; nothing downstream that expects actual weights (a KSampler doing real denoising, for example) will work with it. Its only legitimate destination is a scheduler node that only cares about the sigma math.