LoraLoaderModelOnlyExtractedPromptLite

LoraLoaderModelOnlyExtractedPromptLite is the stripped-down sibling of LoraLoaderExtractedPromptLite: same prompt-tag extraction, same automatic LoRA loading, but no clip involvement at all. You feed it a model and a prompt string, it applies any <lora:name:strength> tags it finds to the diffusion model only, and it returns the patched model plus the tag-free prompt string.

The difference matters. Normal LoRA loading patches both the model and the CLIP text encoder, because many LoRAs were trained with a style component in the text conditioning. A model-only load applies the LoRA's weights to the model and leaves the clip untouched - in the code, the clip-side strength is set to zero. That's what you want for LoRAs that are purely visual, for when you're deliberately keeping the text encoder clean, or when you're reusing a clip that shouldn't change between branches of your graph.

The trade-off deserves honesty: with no clip input, the pack can't apply the text-encoder half of a style LoRA, so some LoRAs will look weaker or different than a full load. If a LoRA ships with trigger words and expects its own prompt style, a model-only load can underdeliver. Use it when you know the LoRA is model-focused - not as your default for everything.

Everything else matches its sibling: same <lora:name:strength> syntax, same lbw/block-weighting extras, same R/U random strength values, same tag-stripped string output. If you only remember one thing, remember that this node has no clip socket - that's the entire difference. In a preset-driven workflow you'd pick it when the clip is already handled upstream and you just want the model patched before sampling.

Install via ComfyUI Manager (search "ComfyUI_mittimiLoadPresetLite") or git clone https://github.com/mittimi/ComfyUI_mittimiLoadPresetLite into custom_nodes, then restart. It needs nothing beyond the pack's single toml dependency, and reads LoRAs straight from your existing loras folder.