LC Sampler Configure (pipe)

Naming alert: this node is called "LC Sampler Configure (pipe)" and its class is LCSamplerConfigurePipeOut, while its unpacker sibling is class LCSamplerConfigurePipe displayed as "Pipe Out". The pack crossed the names over. What you need to remember is this one: it has the widgets. It's the full dual-pass sampler config (steps, swap, both CFGs, sampler, scheduler) with an optional pipe in on the left and a pipe out on top.

So if you're building a pipe-based graph and you want a single node that both configures the dual-pass schedule and packs those values into the bundle, this is the one. You set the dials like you would on plain LC Sampler Configure, and the values get written into the LC_PIPE; any other keys already in the pipe pass through untouched.

How it works

The mechanism is "widgets overwrite pipe, pipe keeps the rest." You can:

• Feed it a pipe (optional) - the sampler keys from this node's widgets overwrite matching keys in the incoming pipe; every other field (model, clip, conditioning, whatever the bundle carries) passes through.
• Or start fresh - leave the pipe input empty and it packs a new bundle from the widget values alone.

Either way it emits the pipe plus the same eight sockets as the non-pipe version: total_steps, cfg_1, denoise, step_swap, cfg_2, sampler_name, scheduler, detailer_steps. So you can chain the pipe onward for the rest of the graph and wire the individual values straight into a sampler if you want.

The dials you actually set

• total_steps and step_swap - the full schedule length and the first→second pass handoff index.
• denoise - 1.0 = full schedule.
• cfg_1 / cfg_2 - guidance for the high-sigma and low-sigma passes. Remember the model family decides what's sane: 5–9 on SD 1.5/SDXL, ~1 on guidance-distilled models (Klein, Z-Image Turbo). CFG 2 is the default for the second pass, which is fine on some architectures and quietly wrong on others.
• sampler_name / scheduler - the standard dropdowns (44 samplers, 9 schedulers). Flow-matching wants Euler-family + conservative schedules; Karras is an SDXL-era habit that breaks flow-matching models.

Install

LC123 pack - Manager (search "LC123") or:

cd ComfyUI/custom_nodes
git clone https://github.com/lonecatone23/ComfyUI_LC123_nodes

Restart. No extra pip packages; cloud ComfyUI hosts (RunningHub etc.) need LC123 requested or installed locally before these nodes resolve.

If you never touch pipes, use plain LC Sampler Configure instead - this one only earns its keep in pipe-based graphs.