WanVideo UltraVico Settings

Straight up: this is one of the newest, least-battle-tested nodes in the whole wrapper, and I'd rather tell you that than pretend otherwise. Its own description points at thu-ml/DiT-Extrapolation, a research project on getting diffusion transformers to generate well past what they were trained on - more frames, higher resolution - without quality falling off a cliff the moment you go beyond the training window. UltraVico looks to be Kijai wiring that idea into Wan. Whether it holds up in the wild is genuinely unproven right now; there's no meaningful community track record on it yet, so treat this as a bleeding-edge knob rather than a settled feature.

How it works, as best the schema tells us

This node patches your Wan model with UltraVico's extrapolation behavior and hands the patched model back - the same shape as other "settings" nodes in this pack (TeaCache, sampler settings) that tweak how the model behaves without touching its weights. Critically, the node's own description says the patch isn't enough on its own: you also have to separately set your attention mode to sageattn_ultravico (wherever your graph configures attention - the model loader or an attention-override node) for any of this to actually do anything. Skip that step and this node's settings are inert.

The inputs and output that matter

Just two required inputs:

• model (WANVIDEOMODEL) - your loaded Wan model, going in to be patched.
• alpha (default 0.9, range 0–1, step 0.001) - the only tunable here. Read it as a blend or strength setting: how strongly the extrapolation behavior is applied. The default sitting near the top of the range (0.9) suggests it's meant to be applied close to fully by default, with lower values presumably pulling generation back toward standard behavior.

Output is a single model (WANVIDEOMODEL) - the patched model, which you then feed into your sampler exactly like an unpatched one.

How to install it

Ships with the pack, no separate download. ComfyUI Manager: search ComfyUI-WanVideoWrapper, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper
pip install -r ComfyUI-WanVideoWrapper/requirements.txt

then restart. The real dependency isn't a model file - it's whatever the sageattn_ultravico attention mode requires under the hood. Because this is a fresh addition, check the pack's example workflows and its requirements.txt for the exact build it expects rather than assuming your existing SageAttention install covers it; extrapolation-aware attention kernels are typically a separate build from the standard one.

Common issues & troubleshooting

I don't have real-world failure reports to hand you here, and I'd rather say that plainly than invent plausible-sounding ones. What I can tell you from the schema and the node's own description:

Nothing happens even with alpha set. This is the one documented gotcha, straight from the pack: the attention mode has to be sageattn_ultravico elsewhere in your graph, or this node's patch has nothing to act on. Check that first before assuming the node is broken.

You're not sure you need this at all. If you're generating within Wan's normal 81-frame, standard-resolution range, you almost certainly don't - extrapolation techniques exist specifically for pushing past those limits, and the standard path is better-understood and better-supported. Reach for this only if you're deliberately trying to go beyond what plain Wan comfortably handles, and go in expecting to do some of your own troubleshooting.