DIGIT LoRA Loader (Model Only)

Most LoRA loaders apply the adapter to two things: the diffusion model and the CLIP text encoder. Sometimes you only want one of those. This node is the model-only half of the DIGIT LoRA Loader family - it applies the LoRA to model, leaves the clip completely alone, and still hands you the auto-detected trigger words. If a LoRA is dragging your prompt interpretation off-course, or you're deliberately keeping the encoder stock, this is the loader to reach for.

The split exists because not every LoRA needs both halves, and on newer LLM-encoder architectures (Flux, Qwen-based models) the encoder isn't even trained as part of a LoRA run - so forcing clip modification can be pure downside. A model-only load keeps your text conditioning untouched while still getting the adapter's full effect on the denoiser.

How it works

Same mechanism as its full sibling: it reads just the metadata header of the .safetensors file (no weight loading), extracts the trigger word using the priority order that covers AI Toolkit, Kohya/sd-scripts, SimpleTuner, and modelspec formats, then applies the LoRA to the model tensor - and only the model. The CLIP input simply doesn't exist on this node, so there's no clip to misconfigure.

The inputs:

• model - the diffusion model to modify.
• lora_name - dropdown from your loras folder.
• strength_model - how strongly the adapter modifies the model. Default 1.0.
• lora_path_override - absolute path to a .safetensors outside the loras folder (e.g. on a network share). Overrides lora_name.
• trigger_words_override - manual trigger words that beat auto-detection.

Outputs:

• model - the modified model, ready for the sampler.
• trigger_words - the detected trigger, ready to wire into a Prompt Combine.
• metadata - training metadata summary from the safetensors header.
• lora_info - a one-line summary (name, base model, rank, trigger).

Note the difference from the full loader: no clip output, and no strength_clip input. That's the entire point - nothing about the text encoder changes.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/thedepartmentofexternalservices/comfyui-digit.git
cd comfyui-digit
pip install -r requirements.txt

(Or ComfyUI Manager → search comfyui-digit → install.) Restart ComfyUI, look under DIGIT.

Common issues

Same metadata story as the full loader: a LoRA with a stripped header returns an empty trigger_words, and the fix is trigger_words_override. And the same base-model caution applies - a LoRA trained for Flux won't help an SDXL checkpoint, no matter how the loader applies it.

The one thing to actually decide before you use this: do you want the clip modified or not? For SD1.5/SDXL-era LoRAs the answer is usually "both" (that's how they were trained). For Flux/Qwen-era LoRAs, model-only is often the better call because those runs never touched the encoder. When in doubt, the full DigitLoraLoader with a modest strength_clip is the safer default.