LC Model Optimizer Video (pipe) ⚡

LC Model Optimizer Video (pipe) is the video loader with an LC_PIPE added in front: same benchmark-driven file picking, same speed-up panel, same six roles - model, clip, video VAE, audio VAE, latent upscaler and LoRA - plus one bundled wire carrying the whole generation's state.

What the pipe adds

The LC_PIPE type is LC123's context object: model 1 / clip 1 / VAE 1, model 2 / clip 2 / VAE 2, image, mask, width, height, latent, batch, positive and negative prompt with their conditioning, seed, total_steps, cfg_1, denoise, step_swap, cfg_2, sampler name, scheduler and detailer steps. On the video version, slot vae_2 is where the audio VAE lives, which is a neat trick: a bundle designed for two-stage sampling carries a second VAE anyway, and a video model with native stereo audio needs exactly that.

Alongside pipe you still get model, clip, vae, audio_vae, latent_upscaler and summary as normal sockets, so adopting the pipe on a video graph doesn't force you to rewrite it today. Wire pipe into LC Sampler Configure Pipe Out when you want the sampler settings to arrive from the same object that carried the weights, or into LC Pipe Combine when you want to override just one slot - a different length, a swapped latent - from a second branch.

Sibling nodes in the same corner of the pack: LC MiniMax H3 Pipe and LC MiniMax H3 Pipe V2 handle the H3-specific payload (references, size, frame rate, prompt, sampler settings) if you're building H3 graphs rather than LTX ones.

The trade, stated plainly

Pipes exist to kill spaghetti, and they do. One wire instead of six also means you can no longer see what's inside it. Unpack model off a pipe that was assembled before your LoRA patch and you'll run base weights with a wire that looks perfectly connected. That's the failure mode of the whole pattern, not of this node: with six wires you can see the LoRA isn't feeding the sampler, with one pipe the connection is there and the contents are wrong.

So: pipe where the graph has settled, explicit wires where you're still experimenting. And remember the pipes don't interoperate - an LC_PIPE isn't rgthree's Context or ComfyUI-Easy-Use's pipe dict, so a workflow that adopts one pack's bus is married to that pack.

Install

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

Restart, or install from ComfyUI Manager by searching ComfyUI LC123 Nodes. No pip step - no requirements.txt in the pack at all. Add ComfyUI-GGUF if you want the quantised weight options.

Troubleshooting

Stale pipe. The number one cause of "it loaded but the model is wrong". Check where the pipe was packed relative to the patch or LoRA node that was supposed to modify it.

Gated, geofenced downloads. MiniMax H3's licence excludes the EU, UK, Korea and the US, and both video families ship gated repos. Accept the licence on Hugging Face and log in locally, or the first run dies mid-queue.

Cross-version pipes. Slots that didn't exist when a workflow was saved simply arrive empty rather than erroring, so a pipe from an older workflow may quietly be missing what a new node expects.

The 1.0-vs-1.1 trap on LTX 2.3. If you're supplying your own distilled LoRA instead of the picker's, use 1.1. Version 1.0 produced garbled audio out of the first sampler stage, which is a miserable thing to debug from the output end.