CWK Wan2.2 Pipeline Splitter

Not every multi-clip Wan 2.2 video needs the pack's loop machinery. Sometimes you just want to build five clips statically - maybe to see all of them at once, or because you're feeding separate video-combine sections - and for that you'd rather reach for the pipeline's dumb cousin: the Pipeline Splitter. It takes the WAN22_PIPELINE object from the pack's Prompt Composer plus a clip_index, and unpacks that one clip into plain, wireable values: prompt, frame_count, lora_stack_high, lora_stack_low, and seed.

That's the entire job. The Composer hands you a whole story in one opaque object; this node is how you open it at a specific page.

Why you'd use it instead of the loop

The Loop Open node drives the same data through re-queues, which is right when you want one sampler section to walk through every clip. But that means waiting on the loop and re-queuing ComfyUI repeatedly. The Splitter is for the opposite instinct: duplicate your KSampler section once per clip, hard-wire each section to its own clip_index, and let the whole multi-clip render happen in a single queue. No state, no re-queues, no surprise restarts. You trade graph size for predictability, which is a fair trade if your clips are few and your VRAM is fine with a big graph.

It also earns its keep as a debugging tool - dropping one Splitter on the Composer's output and watching a single clip's prompt and seed is the fastest way to confirm your blocks are what you think they are before committing to a long render.

Inputs and the gotcha

• pipeline - the WAN22_PIPELINE. Required.
• clip_index - 1-based (min 1, max 64), and this is where people get burned: an out-of-range index doesn't error. The backend logs a message and returns empty outputs - empty prompt, frame_count 1, empty LoRA stacks, seed 0. Your graph will happily run with a blank clip and you'll only notice when that section of the video is nonsense. If a section comes out blank, check your index against the number of blocks in the Composer.

The outputs wire straight into the same places the loop's do: prompt and frame_count into your sampling section, the two LORA_STACKs into the pack's LoRA Applier (high-noise stack to the high-noise model, low-noise stack to the low-noise model - Wan 2.2's two-pass split), and seed into a seed input.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/cowneko/CWK_Wan2.2_Nodes.git

or ComfyUI Manager → CWK Wan2.2 Nodes, then restart. No extra Python dependencies - this is a pure backend node, no custom UI.

It's the least glamorous node in the pack and there's nothing clever to say about it. That's kind of the point: when you want the Composer's structure without the loop's moving parts, this is the boring, reliable way to get it.