Split Loader (full-pipe)

Modern checkpoints stopped being single files. The newer model families - the Qwen-based image models like Anima, Flux-style split checkpoints - ship as a separate diffusion model, a separate text encoder, and a separate VAE, and the classic "load checkpoint" node has nothing to say to them. Split Loader (full-pipe) is this pack's answer: three dropdowns for unet_name, clip_name, and vae_name, and it assembles them into a ready-to-use FULL_PIPE.

It's the node you want if you're running the pack's Anima workflow or anything that gives you diffusion_models/, text_encoders/, and vae/ folders instead of one .safetensors. The pack's own Anima example workflow is built on exactly this - anima-base-v1.0.safetensors as the UNet, a Qwen text encoder, and the Qwen image VAE, wired in as separate files.

The inputs that matter

• unet_name - the diffusion model, from ComfyUI/models/diffusion_models/.
• clip_name - the text encoder, from ComfyUI/models/text_encoders/.
• vae_name - from ComfyUI/models/vae/.
• stop_at_clip_layer - CLIP skip (0 disables). Different from the pack's checkpoint loader, which defaults to -2; here the default is 0.
• shift - the AuraFlow sampling shift, 0 to 10. The pack applies ModelSamplingAuraFlow with this value, which matters for AuraFlow-family models; leave 0 unless your model's docs tell you otherwise.
• seed - dropped into the pipe.
• Optional: lora_stack (apply a stack of LoRAs at load time - same mechanism as the pack's checkpoint loader, based on Easy-Use's loraStack) and apply_negpip (patches CLIP for negative-weight prompts; requires pamparamm/ComfyUI-ppm).

Output is a single full_pipe with model, clip, vae, seed, and meta (stashing the unet name).

The caching detail

Both loaders in this pack keep a one-entry cache keyed by everything that affects the loaded model - names, CLIP skip, shift, the LoRA stack. Re-running a workflow with the same inputs is a cache hit and skips reloading entirely, which is a big deal with multi-GB split checkpoints. Change any of the keyed inputs and the cache is cleared and reloaded. If you're swapping models in a long session, this is what keeps it fast.

Installing it

ComfyUI Manager (search comfyui-mudknight-utils) or:

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

No Impact Pack needed for this one - it's pure ComfyUI core calls. apply_negpip is the only optional dependency (ComfyUI-ppm), and you can ignore it.

Where people get burned

The dropdowns only list what's in the right folders - put a split checkpoint in checkpoints/ and it won't appear. And the shift input is AuraFlow-specific: cranking it on an SDXL-class model is wrong, and leaving it at 0 on an AuraFlow model that wants a shift will give you washed-out results. Match the shift to whatever the model card says. Also worth knowing: unlike a checkpoint loader, nothing here guesses model type - you're responsible for picking a compatible UNet/CLIP/VAE trio, and an encoder from the wrong architecture will fail at encode time, not load time.