TeaCache For Img Gen

This is the node TeaCache started as, before the pack's maintainer folded everything into one. If your workflow - or a tutorial you're following - still calls out "TeaCache For Img Gen" by name, you're looking at the Flux-only predecessor of the general-purpose TeaCache node: same caching trick, same math, just scoped to one model family and missing a couple of the newer controls. Unless you're deliberately reproducing an old workflow, use TeaCache instead - it does everything this node does, plus more.

How it works

The mechanism is identical to its successor: TeaCache estimates how much a denoising step's output would differ from the cached one, and skips recomputing it when that estimate falls under your threshold. It's training-free - a runtime patch on the model, not a fine-tune - which is why it slots in as a model-in, model-out node with nothing to download.

The inputs and outputs that matter

The interface is deliberately narrow. model_type has exactly one option, flux, because that's all this node ever supported - there's no real decision to make there. rel_l1_thresh (default 0.4) is the only dial that matters: push it down if quality suffers, up if you want more speed and can tolerate the softening. Notice what's missing compared to the unified node - no start_percent/end_percent to restrict caching to part of the schedule, and no cache_device choice. Those got added when the pack's options were redesigned in May 2025, well after this node had already been superseded. Output is a patched model, same as always - wire it into your sampler.

How to install it

Same pack as the unified node. Via ComfyUI Manager: search "ComfyUI-TeaCache", install, restart. Manually:

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

No model files to download - this is a code-only node that patches a Flux model you've already loaded.

Common issues & troubleshooting

You inherited a workflow using this node and it errors after updating the pack. That's expected wear on an old node path. Swap in TeaCache with model_type set to flux and you'll get the same behavior plus the options this node never had.

You want to cache only part of the schedule. You can't, here - that control simply doesn't exist on this node. Move to TeaCache for start_percent/end_percent.

Speed isn't matching what the README promises. Double check you're actually running base Flux, not FLUX-Kontext or something adjacent - this node was only ever tuned against base Flux, so the recommended rel_l1_thresh of 0.4 assumes that specifically.

Multiple TeaCache-family nodes in one workflow, and the threshold seems to follow only one of them. That was a real bug early in this pack's life - with more than one TeaCache node present, rel_l1_thresh would silently use whichever node's value was set last, not each node's own setting. It's long since fixed, but if you're running a genuinely ancient pinned copy of the pack, update it before you spend an hour convinced your graph is wired wrong.