🅛🅣🅧 MotionCache Model Patch

If you have a working LTX-2.3 workflow that goes MODEL -> Guider -> SamplerCustomAdvanced, the last thing you want is to rip out the sampler and rebuild around a replacement node. This one doesn't ask you to. Ltx23MotionCacheModelPatch is a MODEL -> MODEL patch that layers caching acceleration on top of whatever sampler you already use - the safest option in this pack, and the one the included workflow demonstrates.

It's also the only node in the pack that doesn't need the LTX Python runtime at all. It delegates to ComfyUI's own built-in EasyCache wrappers (comfy_extras.nodes_easycache), which ship with ComfyUI. No ltx_core imports, no LTX2_REPO_ROOT. If you just want caching on your existing LTX pipeline without the setup drama, this is the node.

How it works

In the source, the node clones your MODEL, drops an EasyCacheHolder into its transformer_options, and attaches ComfyUI's standard EasyCache wrappers - the outer-sample, calc-condition-batch, and diffusion-model wrappers. That's the same machinery as ComfyUI's built-in EasyCache node, so the behavior and its caveats are the well-known ones.

EasyCache's trick: when the denoising input barely changes between steps, it reuses the previous step's cached activations instead of recomputing them. That's a real speedup on video workflows, because most tokens genuinely do change only a little from step to step. But it's not free. Community reports are consistent: speedups vary a lot, and quality takes a hit on fast, complex, or realistic motion - the classic recommendation is to use it on low-motion shots like backgrounds and scenery, not on a dancer. Treat it as a draft-mode accelerator, not a free lunch. (For reference, the other well-known LTX accelerators - NAG from Lightricks, MagCache - sit in the same category: real but quality-sensitive.)

Inputs and outputs

Inputs:

• model (MODEL) - your LTX-2.3 model, after LoRAs and any NAG. The README's recommended chain is LTX model / LoRA / NAG -> Ltx23MotionCacheModelPatch -> Guider / SamplerCustomAdvanced.
• reuse_threshold (default 0.20) - how aggressively to reuse cached activations. Higher = more aggressive = faster but riskier.
• start_percent / end_percent (defaults 0.15 / 0.95) - the sampling-progress window where caching is active. The default skips the very start (where the image is still forming) and the very end (where fine detail settles). Those defaults are sane; leave them.
• subsample_factor (default 8) - the subsampling factor the cache wrapper uses.
• verbose (default false) - prints cache debug details to the console when you're trying to see what it's doing.

Output: a single model (MODEL) - wire it into your Guider exactly where the original model went. That's the whole drop-in story.

Installing

starsFriday/ComfyUI-LTX23-MotionCache, via ComfyUI Manager (search "LTX23 MotionCache") or:

cd ComfyUI/custom_nodes
git clone https://github.com/starsFriday/ComfyUI-LTX23-MotionCache

Restart ComfyUI. No pip dependencies - and unlike the two standalone sampler nodes, this one needs no LTX source tree importable, because it uses ComfyUI's native EasyCache. The pack still expects you to have a working LTX-2.3 environment (weights, the official ComfyUI-LTXVideo nodes); it just doesn't need them as Python imports. The bundled workflow (workflow/LTX-2.3-MotionCache.json) runs this exact path - import it, fix any missing-node warnings (external LTX/video-helper/upscaler nodes must be installed separately), and adjust model names, prompts, and resolution to your setup.

Common issues

• Quality drops on moving subjects - that's EasyCache doing its thing, not a bug. Lower reuse_threshold, or skip caching for shots with real motion and reach for it on static backgrounds.
• No speedup at all - some setups get little from EasyCache; the savings depend on how much the input actually changes per step. Try raising reuse_threshold, and check the console with verbose on to see whether the cache is engaging.
• Conflict with other acceleration nodes - don't stack this with another EasyCache/MagCache wrapper on the same model; you'll get double wrapping and unclear results. Pick one accelerator per model.