D2 Load Lora

The standard LoraLoader takes one LoRA and one strength. D2 Load Lora takes a text block, so you can list several LoRAs with their strengths in plain text and have them all applied at once - and because the spec is just text, you can drive it from a string node, a preset, or an XY plot sweep instead of hand-picking files. If you test a lot of LoRAs, this is the loader that turns "swap a LoRA" into "change a word in a text field."

It has two modes, and the difference matters:

• simple - one LoRA per line, lora_name:model_weight:clip_weight. Leave a weight off and it defaults to 1; give just one weight and both model and clip get it. Multiple LoRAs on one line separate with commas, and lines starting with // or # are comments.
• a1111 - the Stable Diffusion webui style, written inline in a prompt: <lora:lora_name:1>. The node scans the text, applies each <lora:...> it finds, and hands you the remaining prompt with the tags stripped out as formatted_prompt.

That A1111 mode is the same syntax the D2 KSampler understands in its own positive prompt field - so this node is really the explicit, inspectable version of the same feature.

The inputs

• model / clip - the base model to apply to (wire from a checkpoint loader).
• prompt - the multiline text block holding your LoRA specs (plus any prompt, in a1111 mode).
• mode - simple or a1111.
• insert_lora - a dropdown listing your loras; handy for inserting a filename into the text without typos.

Outputs

MODEL / CLIP with everything applied, the raw prompt passed through, and formatted_prompt (the a1111 mode's cleaned-up text, which you'd feed to a text encoder).

Installing it

ComfyUI Manager → search "D2-nodes-ComfyUI" → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/da2el-ai/D2-nodes-ComfyUI

Restart. No downloads from the pack itself - your LoRA files must already live in models/loras. Deps are piexif and charset-normalizer. Since v32.0.0 the pack needs a recent ComfyUI (V3 schema).

The gotchas

Weight semantics are easy to misread: foo:0.8 sets both model and clip strength to 0.8, not just model - if you want clip weaker, write foo:0.8:0.5. And a LoRA with a strength you didn't intend (defaults silently to 1) is the most common "why is this overdone" culprit. Also, in a1111 mode the node applies every <lora:...> in the text - so if you paste a prompt that still contains a tag you meant to keep as text, it'll be applied. Clean, tested text in, predictable strengths out - that's the whole game with this node.