TeaCache For CogVideoX

Different node family, same idea, one big catch: this one doesn't plug into a normal ComfyUI model. Its model input and output are typed COGVIDEOMODEL, not the standard MODEL every other node in this pack uses - which means it won't connect to Load Diffusion Model at all. The README says it outright: this node needs kijai's ComfyUI-CogVideoXWrapper nodes to load and run CogVideoX in the first place. TeaCache here is a companion to that wrapper, not a node you can drop into any graph.

Where it fits

Worth knowing before you build around this: CogVideoX (THUDM/Alibaba) was one of the first video models to prove that open local video generation was viable at all, and it earned that reputation before Wan 2.1 arrived and took over the space. If you're starting a video workflow from scratch today, Wan is almost certainly the better use of your time - and TeaCache's main node already covers Wan directly. This node earns its keep specifically if you've already got a CogVideoX pipeline running through kijai's wrapper and want to speed it up without rebuilding around Wan.

The inputs and outputs that matter

The interface is the simplest in the pack, because it doesn't need to know which model family it's patching - it only ever patches CogVideoX. enable_teacache is a plain on/off switch (default on): flip it off to A/B against the uncached baseline without pulling the node out of the graph. rel_l1_thresh (default 0.3) is the only tuning knob - the README's own number for CogVideoX is 0.3 for roughly a 2x speedup, in line with what TeaCache gets on comparable models elsewhere in the pack. Output is a patched COGVIDEOMODEL - feed it straight into the CogVideoX sampler node from kijai's wrapper pack, not a native KSampler.

How to install it

Two packs, not one. This node ships inside ComfyUI-TeaCache - ComfyUI Manager search "ComfyUI-TeaCache", or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/welltop-cn/ComfyUI-TeaCache
cd ComfyUI-TeaCache
pip install -r requirements.txt

You also need ComfyUI-CogVideoXWrapper installed separately (search for it in Manager, or clone it the same way) - without it you have nothing to plug this node into.

Common issues & troubleshooting

It won't connect to my model loader. You're using a native or unrelated loader that outputs MODEL. This node only accepts COGVIDEOMODEL, which comes exclusively from kijai's CogVideoXWrapper loader nodes - install that pack and load your CogVideoX checkpoint through it.

Quality drops noticeably at default settings. Bring rel_l1_thresh down from 0.3 the same way you would on the main TeaCache node - it's identical threshold logic, just scoped to one architecture.

You're not sure this is even the model you should be running. Fair question. CogVideoX predates Wan and was superseded once Wan 2.1 landed with better quality and a much bigger ecosystem around it. If you're not already committed to a CogVideoX workflow for a specific reason, it's worth checking whether Wan plus the main TeaCache node gets you further with less friction.