DvD LoRA Loader (Model Only)

The DvD LoRA Loader (Model Only) is the stripped-down sibling of the pack's Standard loader: same trigger-word machinery, but no CLIP connection at all. MODEL goes in, MODEL comes out, and the LoRA is applied only to the model weights. It exists for the workflows where the CLIP side of a LoRA is either unwanted or handled elsewhere.

When would you reach for that? The classic case is FLUX and SD3-era pipelines, where the text encoder isn't patched the way it was in the SD1.5/SDXL days and you often load CLIP separately anyway. It's also useful when a LoRA's CLIP tensors are causing trouble and you want to isolate just the model-side patch - or when you simply don't care about the trigger text and want a leaner node. Fewer inputs, fewer wires, less to think about.

Same trigger brains, minus a wire

Everything you'd want from the read side is here. The node reads trigger words in the pack's standard priority order - local .json/.txt sidecar, then the safetensors' internal metadata, then a Civitai hash lookup that writes the JSON sidecar for you. Outputs are MODEL, trigger_text (STRING, comma-joined with anything in pre_text), and info_log. Wire trigger_text into your CLIP Text Encode exactly as you would with the Standard loader.

Under the hood the difference is one line: it calls ComfyUI's LoRA application with the CLIP argument set to None. That's the whole "Model Only" promise - the model gets patched, the text encoder doesn't. Worth knowing if you run a LoRA that stores meaningful text-encoder weights (some character and embedding-style LoRAs do): applied model-only, you may see a weaker or slightly different effect than the model page implies. For the majority of FLUX-style adapters, that's a non-issue.

Inputs and mode

The handful of inputs that matter: lora_name, strength_model (default 1.0, and it's the only strength), mode (Read (只读) / Save (保存修改)), edit_text for save mode, plus the optional pre_text string. Save behavior matches the Standard loader: set mode to Save, type your words into edit_text, queue a prompt, and the node writes an A1111 WebUI-compatible JSON next to the .safetensors, preserving any existing fields.

Install

Same pack, same clean story:

cd ComfyUI/custom_nodes/
git clone https://github.com/idvdii/ComfyUI-DvD-LoraTriggerwordsManager.git
# restart ComfyUI

ComfyUI Manager works too - search "DvD LoRA". Zero extra dependencies, no API key, nothing to configure. If you're already running the pack for its trigger-word handling, this node is the version you slot in when CLIP isn't part of your graph.