DIGIT LoRA Loader

Every LoRA you load has a trigger word baked into its metadata, and ComfyUI's stock loader ignores it. You end up digging through the training config or a README to find out whether it's ohwx or ch9ractername. This loader reads the safetensors header and hands you the trigger words on an output - no extra metadata file, no guesswork. That single feature is why it's worth using even if you never touch the rest of the DIGIT pack.

It's the inference-side loader in a two-node story. This one (DigitLoraLoader) applies the LoRA to your model and clip and returns the trigger text. Its sibling DigitLoraLoaderModelOnly does the same but skips the clip. The pack also ships a different node also displayed as "DIGIT LoRA Loader" (DigitLoRALoader) that's the training-suite companion - it only resolves paths and metadata without applying anything. If you want the adapter actually applied, this is the one.

How it works

The node reads just the metadata header from the .safetensors file - no weight loading, so it's fast - and pulls the trigger word out using a priority list that covers the current training ecosystem: modelspec.trigger_phrase (SimpleTuner, some AI Toolkit versions), then ss_tag_frequency (AI Toolkit stores the trigger as the highest-frequency tag), then ss_training_comment if it mentions "trigger". The pack explicitly supports AI Toolkit, Kohya/sd-scripts, SimpleTuner, and modelspec formats, which covers essentially everything people train with in 2026.

The inputs:

• model - the diffusion model the LoRA applies to.
• clip - the CLIP model, also modified.
• lora_name - dropdown of your loras folder.
• strength_model / strength_clip - both default 1.0. The classic knobs; lower the model strength when the LoRA is overpowering the base.
• lora_path_override - absolute path to a .safetensors outside the loras folder (LucidLink or a network share). Overrides lora_name when set.
• trigger_words_override - type them yourself and it beats the auto-detection.

Outputs:

• model, clip - the modified models, ready for the sampler.
• trigger_words - the auto-detected trigger. Wire it into a Prompt Combine so your prompt literally contains the word the model was trained on.
• metadata - a summary of the training metadata from the safetensors header.
• lora_info - a one-line summary: name, base model, rank, trigger.

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

The trap people actually hit: a LoRA whose metadata is missing or stripped. If trigger_words comes back empty, the file's header just doesn't carry the info - use trigger_words_override and move on. There's no failure you can't recover from here because the loader never fails to load; the worst case is an empty trigger field. And the reminder that applies to every LoRA: the base model matters. A Flux LoRA won't work on an SDXL checkpoint - this node doesn't check compatibility for you, it just applies what you point it at.