VideoForwardOverrider

Video models are where this pack earns its keep. Wan 2.1 takes minutes per clip on consumer hardware, HunyuanVideo wanted 45GB+ of VRAM at launch, and LTX is fast but still pays for every step. TeaCache and First Block Cache skip whole transformer blocks when consecutive steps aren't changing much, and that's worth a lot more on a 14B video model than it is on Flux. VideoForwardOverrider is the node that gives those cache patches a handle on video models: it swaps in the rewritten forward pass that exposes the hook points, and restores it after sampling.

The brief for this node says "Support HunYuanVideo," but read the source and the README and it covers the whole video family: HunYuanVideo, LTXVideo, MochiVideo, and WanVideo. The comfy.icu description just underreports.

Wan needs a little extra care, and this node handles it. Wan's stock forward doesn't pass transformer_options through to the patched path, so VideoForwardOverrider also patches the forward method itself to thread that through - that's why Wan works here while a bare TeaCache hookup on Wan sometimes silently does nothing. Different video models also have different forward signatures (LTX and Mochi pass extra args like frame_rate and num_tokens), which is exactly the kind of fiddly detail this overrider exists to hide.

Wiring

Just model → model, wired as:

UNETLoader (video) → ApplyTeaCachePatch → VideoForwardOverrider → guider/KSampler

The cache node and the overrider both land on the same model object, so their order in the chain doesn't matter.

Install

In ComfyUI Manager as ComfyUI_Patches_ll, or:

cd ComfyUI/custom_nodes
git clone https://github.com/lldacing/ComfyUI_Patches_ll.git
# restart ComfyUI

Needs ComfyUI ≥ 0.3.17. No model files, no heavy Python deps (numpy, packaging only). Note this patches whatever video model you already have - it doesn't bundle Hunyuan or Wan weights.

Caveats worth knowing

• MochiVideo is the pack's known problem child: the README says accelerated output quality is poor and can fail outright (black video). If a Mochi render comes back black, this is the first suspect.
• Wan thresholds vary. The author's speed table has dashes for Wan because different Wan models respond very differently; the TeaCache nodes expose wan_coefficients for exactly that reason. Match it to your Wan variant.
• Don't feed it a Flux model. It checks the model type and no-ops on non-video models, which is safe but confusing if you expected a speedup.

For a mixed Flux-plus-video setup, DitForwardOverrider auto-detects and covers both - this node is the video-scoped version you'd use in a video-only graph.