Anymatix Lora Loader

A LoRA is the cheapest way to change what a model does - a style, a character, a specific look - without retraining anything. ComfyUI's stock LoraLoader is the node that bolts one onto your model and CLIP, with strength sliders for each. AnymatixLoraLoader is the version this pack ships, and it differs from stock in two ways that matter.

First, lora_name is a string path rather than a dropdown of models/loras, so it pairs with AnymatixFetcher: the fetcher downloads the LoRA from its URL, hands this node the path, and the graph runs on a machine that had neither the file nor a dropdown entry for it. Second, it caches. Each instance remembers the last LoRA it loaded and skips the disk read on the next run if you feed it the same file - so in a video workflow that evaluates the same node hundreds of times, you're not re-reading tens of megabytes of weights per frame. It's the kind of optimization you only notice when it isn't there.

How it works

The inputs are the familiar LoRA quartet:

• model (MODEL) - the diffusion model the LoRA patches.
• clip (CLIP) - the text encoder it also patches.
• lora_name (STRING) - path (fetcher output works) or, since this node treats the input as a literal path, a real filesystem path to the .safetensors.
• strength_model and strength_clip (FLOAT, default 1.0) - how hard to apply. Both can go negative (inverting a style is a real trick), and if both are 0 the node short-circuits and passes the model straight through - cheap way to A/B a LoRA without deleting it from the graph.

Outputs: MODEL and CLIP, both modified, which continue into your sampler. Chain loaders by feeding one's MODEL/CLIP outputs into the next - multiple LoRAs stack that way, and it's the standard pattern for combining a style LoRA with a character LoRA.

What "path, not dropdown" means here

Pay attention to the difference between this node and the pack's other LoRA loader. AnymatixLoraLoaderModelOnly resolves bare filenames against your loras folder; this one treats lora_name as a literal path and will fail on just a filename. So: full path, or fetcher output, and you're golden.

Install

Same pack story:

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

then restart, or ComfyUI Manager → search "anymatix-comfy-nodes".

Issues to expect

Where people get burned: loading a LoRA trained for a different base model (an SDXL LoRA on an SD 1.5 checkpoint does nothing useful and sometimes actively degrades output), and strength stacking - two LoRAs each at 1.0 can oversaturate a style fast, so start at 0.6-0.8 when combining. If a LoRA "loads fine but has no effect," check the base model match before you blame the node.