Anymatix Lora Loader Model Only

Most LoRAs you'll meet patch two things at once: the diffusion model and the CLIP text encoder, because a style LoRA usually wants its own way of reading words too. But plenty of LoRAs don't need the CLIP half - a pose, motion, or fine-detail LoRA works entirely on the denoising side, and forcing it through the text encoder is just extra work and a chance for weird interactions with your prompt. AnymatixLoraLoaderModelOnly is ComfyUI's LoraLoaderModelOnly with the pack's path-based treatment: it takes a model, a lora_name, and a single strength_model slider, and it only touches the diffusion model.

It's the right node when you want a clean separation - keep your text conditioning untouched while a model-only LoRA steers motion or fine structure. Multiple of these chain together exactly like the full loader: feed each one's MODEL output into the next, and each adds its own patch.

How it differs from AnymatixLoraLoader

Two practical differences worth knowing, because people pick the wrong one:

• It skips CLIP entirely. No clip input, no strength_clip. If your LoRA does have a text-encoder component you want (typical for style/character LoRAs), use the full AnymatixLoraLoader.
• It resolves filenames, not just paths. This one subclasses ComfyUI's own loader and strips the input to a basename before resolving it against models/loras. So a bare filename like my_lora.safetensors works if the file is in your loras folder - no absolute path required. That's a friendlier behavior than the full loader, which takes paths literally.

Inputs and output

• model (MODEL) - the diffusion model to patch.
• lora_name (STRING) - filename or path to the LoRA.
• strength_model (FLOAT, default 1.0) - how strongly to apply. Negative values are allowed.
• Output: MODEL, modified → straight into your sampler or another chained LoRA node.

Install

The usual pack routine:

cd ComfyUI/custom_nodes
git clone https://github.com/Anymatix/anymatix-comfy-nodes

then restart, or install via ComfyUI Manager by searching "anymatix-comfy-nodes". No extra downloads beyond the LoRA file itself.

Issues to expect

The trap is subtle and it's about which LoRA you hand it. Many community LoRAs assume their CLIP component will be applied - the style was baked into both halves during training. Feed one of those to a model-only loader and you'll get a weaker or subtly different result than the creator intended. The rule of thumb: model-only loaders shine for structural LoRAs (pose, motion, detail); if in doubt, use the full loader and set strength_clip lower. Also remember it doesn't download anything - pair it with AnymatixFetcher if the file isn't local yet.