Load LoRA for CLIP Only

Most LoRAs touch both halves of the pipeline: the UNet (or diffusion transformer) and the text encoder. But not all of them need to, and some shouldn't. UC_LoraLoaderCLIPOnly applies a LoRA only to the text encoder - the CLIP - and leaves the diffusion model completely untouched. You get a modified clip output and no patched model to worry about.

When is that what you want? Three cases. First, text-encoder-only LoRAs: some LoRAs are trained exclusively on the TE (prompt-side style or token triggers), and loading them with a normal LoRA loader either fails or silently does nothing to the UNet anyway. Second, TE-LoRA experiments: this is the clean way to test what a LoRA's text-side component contributes on its own, in isolation from the diffusion model. Third, compatibility: the node deliberately mirrors ComfyUI's Core LoraLoader conventions, so it slots into existing workflows as a drop-in for the CLIP path.

How it works

You connect a clip and pick a lora_name from the dropdown (populated from your models/loras folder). strength_clip controls how hard the LoRA hits the text encoder - range is −10 to 10, default 1.0. Negative strengths invert the effect, which is occasionally exactly what you want for removing a style trait. Output: the modified clip, ready to feed your conditioning.

Three details from the source worth knowing:

• LoRAs without text-encoder data can't be used here. The tooltip says it plainly: if a LoRA has no TE tensors, the node rejects it. That's not a bug - it's the node being honest about what it can do.
• Strength 0 short-circuits. Setting strength_clip to 0 returns your original clip unchanged, no loading work at all. Handy for A/B testing.
• There's a loader cache. Repeated loads of the same LoRA reuse the cached patch, which makes batched runs faster.

The pairing rule to remember: this node gives you a CLIP-only load. If you also want the LoRA on the diffusion model, you still need the normal LoraLoader for that half - these are complementary, not competing.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-UtilsCollection

Restart (or ComfyUI Manager → search "ComfyUI-UtilsCollection"). Part of silveroxides/ComfyUI-UtilsCollection, a ~150-node utility pack; deps are opencv-python and typing-extensions. No models to download - the LoRA files come from your existing models/loras folder.

Where people get confused: they load a TE-only LoRA with a standard loader, see no change, and assume the LoRA is broken. If you've got a LoRA that was trained on prompts alone, this is the node that makes it actually express itself. For anyone poking at text-side conditioning - especially with the newer VLM-encoder models - CLIP-only loading is a sharper tool than the all-in-one loaders, and this is the cleanest version of it in the pack.