LoraLoaderExtractedPromptLite

If you've ever wished ComfyUI would just read the <lora:name:strength> tags people put in their prompts like A1111 does, LoraLoaderExtractedPromptLite is that wish granted. It takes a model, a clip, and a prompt string, scans the prompt for LoRA tags, loads every LoRA it finds, and hands back the patched model, the clip, and the prompt with all the tags stripped out.

The workflow is effectively automatic LoRA loading. Write your prompt with <lora:character_style:0.8> inline, feed it in, and the node resolves that name against your loras folder, globs the actual file, and applies it at the tagged strength. The third output - string - is the prompt with the tag removed, so the cleaned text is what actually reaches the text encoder. Wire that into TextToConditioningLite or a plain CLIP Text Encode and the tag never leaks into the encoding.

The syntax it understands goes beyond the basics, and this is where the pack borrows from Inspire Pack conventions. On top of <lora:name:strength> you get :lbw= block weighting and :start=/:stop=/:step= controls - and if you have ComfyUI-Inspire-Pack installed, it will even read that pack's lbw-preset.txt so named presets work as shorthand. There's also a genuinely odd (and fun) feature: the vector values can be R or U, meaning random. R picks a strength between 0 and 3, U between -1.5 and 1.5, and the seed input drives that randomness - so the pack will happily randomize your LoRA strengths every run if you write <lora:name:1:lbw=R>.

My honest take: 95% of people should use the plain <lora:name:0.8> form and ignore lbw entirely. Block weighting is a tuning rabbit hole most models don't need, and the random R/U syntax is a party trick until you're doing serious experimentation. The everyday win is just having your LoRA tags in the prompt where you can read them - and having the preset hub swap the whole prompt, LoRA tags included, when you change your recipe. That's the real reason this node exists in the pack: presets can now carry LoRAs without you adding a loader per file.

One pairing note: if a LoRA doesn't need to touch the text encoder, use the model-only sibling (LoraLoaderModelOnlyExtractedPromptLite) instead - this node patches both model and clip, which is what most character and style LoRAs want.

Install via ComfyUI Manager (search "ComfyUI_mittimiLoadPresetLite") or git clone https://github.com/mittimi/ComfyUI_mittimiLoadPresetLite into custom_nodes, then restart. Only dependency is toml; the lbw-preset reading is optional and only kicks in if Inspire Pack is present.