Loader (full-pipe)

Loader (full-pipe) is the front door of the mudknight full-pipe system: it loads a checkpoint, sets CLIP skip and a seed, optionally applies a stack of LoRAs, and packs the whole thing into a single FULL_PIPE connection. If you've ever looked at an Easy-Use or rgthree workflow and wished the spaghetti of loaders and reroutes would just collapse into one line - that's this.

The full-pipe idea is the point. Instead of dragging three wires (MODEL, CLIP, VAE) plus a seed everywhere, every node in this pack passes one FULL_PIPE connection that carries model, clip, vae, positive, negative, seed, image, latent, and a meta dict. The author built it explicitly as an alternative to the ComfyUI-Easy-Use pipe system, with the complaint that Easy-Use's pipe-in is slow to rewire. If you're building one of these pipe workflows, this node is where it starts.

The inputs that matter

Three required, and only two you'll touch regularly:

• ckpt_name - dropdown of your checkpoints, same as the built-in loader.
• stop_at_clip_layer - CLIP skip, default -2, 0 disables it. Keep -2 for the Pony/Illustrious/NoobAI anime lineage and SD 1.5; on SDXL it usually changes nothing, and on the 2026 LLM-encoder models (Anima's Qwen3, and friends) the parameter has nothing to attach to at all. If you don't know which camp you're in, leave it alone.
• seed - a uint64, default 0. It gets stored in the pipe so every downstream sampler inherits it.

Optional:

• lora_stack - a LORA_STACK input, so you can chain a stack of Load LoRA (full-pipe) style nodes. If you're used to rgthree's Power Lora Loader, this is the same idea, just pipe-native.
• apply_negpip - off by default. Flip it on and the node applies the CLIPNegPip patch from pamparamm's ComfyUI-ppm so you can use negative weights in prompts. If ppm isn't installed it just prints a warning and continues - no crash, but also no negpip.

How it works

It loads the checkpoint with ComfyUI's own load_checkpoint_guess_config, applies the CLIP layer stop, then any LoRAs, then the optional negpip patch, and hands you a pipe with meta = {"ckpt_name": ...} so the Save node knows what model you used. One detail worth knowing: it caches one loaded model at a time and clears the cache when you switch checkpoints to free VRAM/RAM. That's a feature (no silent memory bloat on long sessions), but it means switching checkpoints mid-run pays the load cost again - the same trade every loader makes, just done explicitly.

The output is a single full_pipe. Wire it into Prompt (full-pipe) or Base (full-pipe) next.

Installing

It's part of the mudknight utils pack, so you get it with the rest. ComfyUI Manager: search comfyui-mudknight-utils. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/mudknight/comfyui-mudknight-utils

Then restart ComfyUI. The pack's requirements pull in huggingface_hub, numpy, opencv-python, pillow, openvino, and onnxruntime. No model files to download for this node - your checkpoints are the models.

Troubleshooting

• LoRA not found warnings - the node looks up LoRAs by filename across your whole loras folder, subdirectories included. If a lora_stack entry doesn't resolve, it prints Warning: LoRA '<name>' not found. and keeps going, so a typo just silently drops the LoRA. Check your stack's file names.
• CLIP skip doing nothing - expected on SDXL; not a bug.
• "CLIPNegPip not found" warning - install pamparamm/ComfyUI-ppm or turn apply_negpip off.

Small pack, zero impressions on comfy.icu - you found it through a specific workflow, most likely the author's own SDXL or Anima examples. That's fine; it does one job and does it without ceremony.