Pipe Out WanVideo Setup

What it is

Wan video generation has a sampler settings problem. Unlike image diffusion, where steps and CFG live happily in one KSampler widget, a Wan workflow scatters its sampling parameters across the setup - and a few of them (model_shift, steps_start, steps_stop) aren't even the kind of thing the stock KSampler exposes. Pipe Out WanVideo Setup is the node that pulls those five values back out of the pipe where WanVideo Setup stashed them, so you can wire them into whatever sampler your workflow actually uses.

In plain terms: WanVideo Setup bundles steps, CFG, and the Wan-specific bits into one pipe; this node unbundles them into individual sockets. Setup in, wires out.

How it works

The mechanism is dead simple and that's the point. The WanVideo Setup node builds a Python dict with five keys - steps, cfg, model_shift, steps_start, steps_stop - and this node reads exactly those keys and emits them as typed outputs: three INTs and two FLOATs. Like everything in the Eclipse pipe ecosystem since v4.0.0, it only accepts dict-style pipes. No computation, no magic, just extraction so your sampler can see the values.

Worth knowing what those values mean in a Wan context, because they aren't the usual suspects. model_shift is a Wan 2.1 thing - a shift applied to the timestep schedule, defaulting to 5.0 in the setup node, and it's one of those parameters that silently changes the whole character of a generation if you get it wrong. steps_start and steps_stop are the split-step bounds some Wan samplers use to run part of the denoise schedule with different settings (a stepped or partial schedule), where -1 means "use the default". If your sampler doesn't have a slot for these, they stay unconnected - the node only emits what the pipe contained.

The inputs and outputs that matter

pipe - one required input, the dict pipe from WanVideo Setup (or any pipe carrying those five keys).

Outputs:

steps (INT) - the total step count, usually small for Wan. The 14B model runs happily at 20–30 full-quality steps, and distilled setups (CausVid/LightX2V) drop to 2–8 at CFG 1.0.
cfg (FLOAT) - classifier-free guidance. Wan's distilled models want CFG 1.0; full models run higher.
model_shift (FLOAT) - the timestep shift. 5.0 is the standard Wan default.
steps_start, steps_stop (INT) - the split-step bounds for samplers that support them.

When you'd use it

Any time you want the sampling parameters for a Wan run defined in one place (the setup node) but applied in another (the sampler). That's the whole job. It's also the honest way to see what's in the pipe - if your Wan video comes out looking wrong and you suspect the sampler is eating bad settings, this node is how you check what actually made it through.

Installing it

It's part of ComfyUI_Eclipse, so:

cd ComfyUI/custom_nodes
git clone https://github.com/r-vage/ComfyUI_Eclipse
pip install -r ComfyUI_Eclipse/requirements.txt

then restart ComfyUI. Manager users can just search "Eclipse". Note the pack's own migration history: it was previously RvTools_v2, and v4.0.0 dropped all legacy nodes - old workflows may need the bundled Workflow Migration Tool before they load.

Common issues

The only real failure mode is the one shared across every Eclipse Pipe Out node: feed it a legacy tuple pipe and it throws. The error message is explicit ("expects dict-style pipes only"), and the fix is to rebuild the pipe with the modern setup node. If steps or cfg come out as 0, the pipe was missing the key - check that you're feeding the output of WanVideo Setup, not some other pipe that happens to be lying around your graph.