Register Model as LoRA Hook (MO) 🎭🅐🅓

The model-only cut of ADE_RegisterModelAsLoraHook, the same relationship ADE_RegisterLoraHookModelOnly has to ADE_RegisterLoraHook: same idea, minus CLIP. This pack's "register a whole checkpoint as a hookable LoRA-like influence" feature - the README's "Models as LoRA" line - without the text-encoder side.

The four-way family

It's worth seeing all four register nodes side by side, since they're really one idea with two independent axes - LoRA file vs. whole checkpoint, and CLIP included vs. model-only:

| Node | Source | Touches CLIP | |---|---|---| | ADE_RegisterLoraHook | LoRA file | Yes | | ADE_RegisterLoraHookModelOnly | LoRA file | No | | ADE_RegisterModelAsLoraHook | Whole checkpoint | Yes | | ADE_RegisterModelAsLoraHookModelOnly (this node) | Whole checkpoint | No |

Use this one when you want an entire checkpoint's UNet-side influence - motion, structure, texture - available as a maskable, schedulable hook, and you don't need (or don't want) that checkpoint's text-encoder behavior mixed in.

Inputs and output

Required: model (MODEL, your base), ckpt_name (a dropdown of your checkpoint files - the one being hooked in), strength_model (default 1, range −20 to 20).

Two outputs: MODEL - the hooked version, route your unhooked base model to anything that shouldn't carry this effect - and HOOKS, the actual payoff, feeding into ADE_AttachLoraHookToCLIP or ADE_AttachLoraHookToConditioning to gate where the effect fires.

How to install it

ComfyUI Manager: search AnimateDiff Evolved, confirm the author is Kosinkadink, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/Kosinkadink/ComfyUI-AnimateDiff-Evolved

You need the checkpoint you're hooking in already in your standard checkpoints folder. No AnimateDiff-specific download for this node itself.

Common issues & troubleshooting

ckpt_name dropdown is empty or missing your file. Confirm it's in the normal checkpoints folder and restart ComfyUI - dropdown lists built from a folder scan don't pick up new files without a restart.

Slower graph build, more memory used than a LoRA hook. Expected - this loads a full checkpoint's weights to compute the hook, the same cost as ADE_RegisterModelAsLoraHook, just without the CLIP portion.

Effect is much stronger than expected at default strength. strength_model at 1.0 applies the full weight difference of an entire checkpoint, not a small LoRA - that's a bigger effect than most people expect from a "strength 1" default. Start well below 1.0.

HOOKS output inert. Correct on its own - it needs an attach node downstream (ADE_AttachLoraHookToCLIP/ADE_AttachLoraHookToConditioning) before it does anything. This node is only registration.

Not sure whether you need CLIP included. If the checkpoint's text-encoder behavior isn't something you're relying on, this model-only version is simpler and one input lighter. Reach for ADE_RegisterModelAsLoraHook instead if you specifically want the CLIP side hooked too.