Model Loader Pipe

This is the same model-loading engine as Model Loader [Eclipse], with exactly one difference you care about: instead of model/clip/vae/audio_vae/model_name on separate sockets, everything comes out of a single PIPE output. Nothing about the loading itself changes - same model_type, same name dropdowns, same LoRA slots, same blocks_to_swap. You're just choosing how the result travels.

Think of it as the middle ground in this pack. The Smart Model Loader is the all-in-one that also creates your latent, sets your sampler, and manages templates. The plain Model Loader gives you direct wires for a conventional graph. Model Loader Pipe keeps your hands on the individual model files but commits to the pipe-first philosophy: one wire from here to wherever the model's actually used, with far less canvas sprawl between them.

The PIPE it emits carries model (always), clip and vae when the checkpoint has them baked, plus model_name, is_nunchaku, and - if you enabled the LoRA or CLIP-trim features - lora_names and clip_skip. So a UNet flow works too: load your diffusion model here, add CLIP Loader and VAE Loader separately, and merge everything with the pack's IO Checkpoint Loader, which takes the pipe in and unpacks every field onto individual sockets. Or skip the unpack entirely and feed the pipe straight into Eclipse KSampler (Pipe), which reads the model and VAE out of it itself.

One honest caveat before you adopt pipe-first everywhere: a single context wire is a wonderful tangle-killer, but it also hides which version of each value is inside. If you unpack model from a pipe that was assembled before a LoRA got applied, you'll silently run base weights - the connection is present, the contents are wrong. It's the same trade every context-bus design makes, and it's worth knowing about it before it bites you mid-experiment. Keep the pipe line short and don't be shy about unpacking at the point where you're actively iterating.

Installing

Same as every node in this pack:

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

or search ComfyUI Smart Model Loader in ComfyUI Manager. Restart ComfyUI. Nunchaku and GGUF model types need the respective ComfyUI-Nunchaku and ComfyUI-GGUF extensions; standard checkpoints and UNets don't.

When to reach for it

• You want a single authoritative model source feeding several consumers (KSampler + VAE Decode + metadata) without three copies of the file selected.
• You're building a modular UNet/quantized graph and want one bundle to carry through IO Checkpoint Loader.
• You like the Smart Model Loader's pipe style but find the chip bar and templates overkill for a fixed setup.

It's a variant node, so there's no deep troubleshooting lore: if the pipe comes out empty, the usual suspects are a model_type that doesn't match your file, a missing extension for quantized formats, or a checkpoint with no baked CLIP/VAE when you expected one - which shows up as clip/vae being None inside the pipe rather than on a socket you can see.