Pipe Out WanVideo Setup

Wan video doesn't sample like an image model, and this node exists to carry its unusual settings without a tangle of wires. "Pipe Out WanVideo Setup" unpacks the pipe from "WanVideo Setup [RvTools]" - the RvTools settings node with the Wan-specific defaults baked in - into the five values the Wan pipeline consumes: steps, cfg, model_shift, steps_start, and steps_end. Plus the pipe passthrough.

Here's what those mean and why they're Wan-specific:

• steps (INT) - default 4. Wan runs at low step counts; with speed LoRAs people go even lower. Four steps is the classic 2.x starting point.
• cfg (FLOAT) - default 1.0. This is the telling one: Wan's guidance default is 1.0, which means effectively no CFG for the standard model. The community guidance boost (CFG 3.5+ on the high-noise pass) only appears in expert-split workflows.
• model_shift (FLOAT) - default 5.0. The sigma-shift for the model's sampling schedule; the standard Wan 2.x value.
• steps_start / steps_end (INT) - defaults 2 and 2. These are the expert-split steps for Wan 2.2's two-expert MoE: denoising hands off between the high-noise expert (motion, composition) and the low-noise expert (detail) at these points in the run. That split is why Wan 2.2 pipelines carry two step values where other video models carry one.

Mechanically, this is the same relay as every RvTools Pipe Out: one pipe input, pipe passed through, each field split to its own output via a tuple unpack. The defaults above come straight from the "WanVideo Setup" node's source - steps 4, cfg 1.0, model_shift 5.0, split steps 2/2.

Why bother bundling them? Because those five numbers are the entire "does my Wan workflow work" dial set, and they belong to the sampler regardless of which part of the graph is executing. Bundle once, unpack near the sampler, and you can't have the KSampler and the model-shift node disagreeing about the run.

Installing it

Ships in ComfyUI-RvTools_v2. ComfyUI Manager: search "ComfyUI-RvTools_v2". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/r-vage/ComfyUI-RvTools_v2

Restart ComfyUI. Dependencies (torch, numpy, Pillow, opencv-python, pilgram, pynvml, piexif) install through Manager. Note the pack itself doesn't include the Wan Video Wrapper - you'll need that (or the equivalent Wan nodes) for the actual sampling.

Where people get burned

• These defaults assume the standard Wan 2.x recipe. If you're on a fine-tune or a speed-LoRA setup, cfg 1.0 and 4 steps are starting points, not rules. The KB's rule of thumb: keep the high-noise pass clean, apply speed LoRAs to the low-noise pass.
• Pipe shape is strict. The unpacker wants the five-value WanVideo pipe from "WanVideo Setup". Any other pipe throws a tuple ValueError at runtime.
• Deprecation. The RvTools v2 pack is unmaintained and the README points to ComfyUI_Eclipse. Works as installed; check the successor before building new Wan pipelines.