LC Model Optimizer (pipe) ⚡

LC Model Optimizer (pipe) is the same loader as LC Model Optimizer with an LC_PIPE output bolted on the front. Same file pickers, same measured-on-your-machine recommendations, same speed-up panel. What changes is how it hands the result downstream - and whether you want that depends entirely on how you feel about wires.

What's different

Instead of model / clip / vae and a summary, you also get a pipe of type LC_PIPE, which bundles the whole context of a workflow in one object: model 1, clip 1, VAE 1, then model 2, clip 2, VAE 2, then image, mask, width, height, latent, batch, positive and negative prompt plus their conditioning, seed, total_steps, cfg_1, denoise, step_swap, cfg_2, sampler name, scheduler and detailer steps. Twenty-five slots in total.

The pipe is the packing-layer answer to the same problem rgthree's Context and Efficiency Nodes' pipe solved years ago: a big workflow stops being forty wires and starts being one, so you can carry the entire state of a generation down the graph and unpack it only where you need it. LC's version pairs with LC Pipe Out, LC Pipe (in/edit), LC Sampler Configure (pipe) and LC Pipe Combine.

The regular model, clip and vae outputs are still there, so nothing forces you to commit. On this particular node the second slot set isn't decoration either: model 2 is Ideogram 4's second model, which is why the pipe version is the one you reach for on Ideogram graphs.

Why you'd pick it

Three honest reasons:

1. You're building or reading a large graph. One wire carrying everything is genuinely more legible than six parallel ones, and you can pack the pipe at the top and unpack at the bottom.
2. You want the sampler settings to travel with the model. LC Sampler Configure Pipe Out unpacks an LC_PIPE into sampler sockets plus a latent, so steps, CFG, sampler and scheduler arrive from the same object that carried your weights.
3. You're using the pack's other pipe nodes. Pipes are the pack's connective tissue, and mixing a pipe node with a spaghetti of individual wires gets you the worst of both.

The reason not to: pipes are a one-way door. Every pack invented its own context type, and they don't interoperate - an LC_PIPE is not rgthree's Context and it's not a ComfyUI-Easy-Use pipe. Adopt one and your workflow is married to it.

The failure mode (say it out loud once)

Bundles hide what's inside them. The classic mistake is unpacking model off a pipe assembled before the node that applied your LoRA, silently running base weights because the wire looks connected and the contents are stale. With six wires you can see the LoRA isn't feeding the sampler. With one pipe, you can't - the connection is present and the value is wrong.

Rule of thumb: keep explicit wires in the part of the graph you're actively iterating on, and use the pipe where the graph has settled.

Install

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

Restart, or search ComfyUI LC123 Nodes in ComfyUI Manager. No pip step: the pack skips requirements.txt entirely and uses ComfyUI's own torch/numpy. If the pipe socket shows as missing type, your ComfyUI didn't reload the pack - a restart fixes that, and Ctrl+F5 clears the browser side after an update.

Troubleshooting

Everything in the plain optimizer node's bag of tricks applies here, since it's the same engine: gated Hugging Face downloads need the licence accepted and a local login or the run dies mid-queue, GGUF files need ComfyUI-GGUF installed, and a mismatched text encoder or VAE gets a ⚠️ read from the file header - it warns, it doesn't block.

Pipe-specific: if downstream nodes behave as if a socket is empty, check whether the pipe reaching them was built before the value was set. And if you're feeding an old workflow, note that pipes carry fixed slots - a slot that didn't exist in the version that saved your workflow will simply be empty, not an error.