WanVideo Tea Cache (native)

Let's get the important part out of the way first: this node is deprecated, and it says so itself. The description reads "DEPRECATED, use the native EasyCache or alternative custom node that's up to date instead of this." If you're building a Wan workflow today, reach for EasyCache. This page exists because the node still shows up in old workflows and old tutorials, and you deserve to know what it was and why it's on its way out.

What TeaCache was for

TeaCache is a temporal caching trick for video diffusion. Across the many denoising steps of a generation, a lot of the work is nearly identical from one step to the next - TeaCache notices when a step is close enough to a cached one and skips the redundant compute. On Wan the payoff is real: it's one of the standard acceleration techniques people stack alongside SageAttention and TorchCompile to get a 4090 from unusable to tolerable. The community's own summary: "smart temporal caching that skips redundant computation across similar frames. Moderate speedup with minor quality impact."

That "minor quality impact" is the catch, and it carries over to whatever cache you use next: people consistently note that disabling TeaCache improves face consistency at the cost of speed. Caching skipped work means occasionally reusing something that should have changed, and faces are where you notice it first.

The inputs and outputs that matter

Model in, patched model out. The settings, for archival completeness:

• rel_l1_thresh (default 0.275) - the similarity threshold for deciding when a step is close enough to skip. Higher skips more aggressively (faster, rougher); lower is conservative.
• start_percent (0.1) and end_percent (1) - the slice of the denoising schedule caching is active over. Early steps set up structure, so caching usually starts a little way in.
• cache_device (offload_device default, or main_device) - where the cache tensors live. Offload keeps them off your GPU.
• coefficients - a preset matched to your Wan variant: 1.3B, 14B, i2v_480, i2v_720, or disabled. Model-specific tuning constants; pick the one that matches what you're running or it misbehaves.

How to install it

It's in kijai's KJNodes pack, but there's no reason to install the pack for this node - install EasyCache instead. If you already have KJNodes for other nodes, this one just lives in the KJNodes/deprecated category.

• ComfyUI Manager - search KJNodes for ComfyUI, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-KJNodes, then pip install -r ComfyUI-KJNodes/requirements.txt, and restart.

Common issues & troubleshooting

It errors after a ComfyUI or Wan update. Don't debug it. The whole point of the deprecation notice is that this implementation is no longer maintained against the moving target that is Wan. Swap in EasyCache (native) and move on.

Faces drift across the clip. That's the caching, not Wan. Whichever cache you land on, turning it off is the first thing to try when identity wobbles - not the last.

Mismatched coefficients. If you point a 14B preset at a 1.3B model (or an i2v preset at a t2v run), the cache's skip decisions are tuned wrong and quality suffers. Match the preset to the model. And if you've inherited an old workflow that uses this node and it still works, you don't have to rip it out this second - but treat it as living on borrowed time.