ZML_LoRA加载器（仅模型）

Some LoRAs only touch the UNet, some workflows only want the model patched and don't want a second CLIP wire, and some people just want a LoRA loader with exactly one knob. ZmlLoraLoaderModelOnly (ZML_LoRA加载器（仅模型）) is that minimal node: lora_名称 dropdown, a 模型 input, a 模型_强度 weight, and a single 输出_模型 output. CLIP is simply not in the picture.

The mechanism is honest and thin. If a model is connected and the LoRA isn't "None", it loads the file with comfy.utils.load_torch_file (safe_load on) and calls comfy.sd.load_lora_for_models(model, None, lora, weight, 0.0) - note the CLIP side gets a 0.0 strength, i.e. "don't touch CLIP, we don't have one." If the model input is empty, it returns your input (or None) untouched. That's the whole node.

When is this actually the right tool? Three cases. First, LoRA packs that only patch the UNet don't need CLIP patching at all, so wiring a CLIP through a standard loader is pointless overhead. Second, in architectures where you're handling CLIP separately (some video and Flux-style pipelines route CLIP through their own path), a model-only loader slots in without fighting your existing CLIP wiring. Third - and this is the ZML-pack reason - the README notes the pack's heavier "power LoRA loader" made CLIP optional after fixing a bug where an unconnected CLIP caused an error. This node is the distilled version of that fix: there's nothing optional to misconnect, because there's no CLIP socket to leave dangling.

The honest limits: no txt/trigger-word output (the five-slot ZmlLoraLoaderFive has that), no multi-LoRA stacking (ZmlLoraListOutput and ZmlLoraLoaderFive cover those), no previews. It's a deliberate "only what it says on the tin" node. And a quiet gotcha from the code: if you leave 模型 disconnected, the function returns the bare input - which in an unconnected state is None - so downstream will fail somewhere else. The node is "model only," not "model optional."

It's also a decent template if you're writing your own LoRA node: the pattern of load_torch_file + load_lora_for_models with a zero CLIP weight is the minimal correct implementation, and this one does it without error-prone extras.

Install is pack-standard - ComfyUI Manager search "ComfyUI-ZML-Image", or:

cd ComfyUI/custom_nodes
git clone https://github.com/zml-w/ComfyUI-ZML-Image

then restart. It sits under ZML 图像 → lora加载器 in the Chinese UI (English patch available from the author's repo). No extra dependencies. For the vast majority of workflows the official LoraLoaderModelOnly-equivalent in core ComfyUI does this job identically - so reach for this one because you're already in the ZML pack and want the same look and behavior across your graph, not because it's magically different.