How to Load LoRA in ComfyUI: A Practical Guide for Direct Path Usage

LoRA (Low-Rank Adaptation) patches offer a lightweight way to customize diffusion models and CLIP encoders for specific styles, characters, or concepts. This guide explains how to load a LoRA in ComfyUI by supplying a raw path and filename, detailing the dynamics between the model and CLIP inputs, the role of strength controls, and practical steps for setup, testing, and advanced usage.

LoRA is a small patch, not a full model: it freezes the base checkpoint and adds low-rank adjustments on top, which is why file sizes run tens to a couple hundred megabytes instead of gigabytes. This node needs both halves of what a LoRA touches, which is why model and clip are separate required inputs - LoRAs commonly modify the diffusion model and the CLIP text encoder differently, so you get independent strength controls for each: strength_model and strength_clip, both defaulting to 1 and both able to go negative if you actually want to subtract a LoRA's effect rather than add it.

Key concepts and defaults

• lora_path and lora_name are split into two fields: lora_path is the directory (defaulting to a Windows-style D:/LORA) and lora_name is the filename within it (defaulting to lora.pt).
• lora.pt is a placeholder filename with a .pt extension, but most LoRAs you download (from CivitAI or elsewhere) ship as .safetensors.
• If you’re using local LoRAs from the standard models/loras folder, the core LoraLoader node provides a friendlier dropdown and avoids path typos-use this dedicated flow when files reside outside the default folder.
• strength_model and strength_clip are floats with a typical range around 0.6-1.0, but can be negative to subtract the LoRA’s influence.

What the LoRA Loader does in ComfyUI

The LoraLoader node is designed to facilitate the loading and application of LoRA models to both the main model and the CLIP model in your AI art generation workflow. This node allows you to enhance and customize the behavior of your models by integrating pre-trained LoRA models, which can significantly improve the quality and specificity of the generated outputs.

Step-by-step usage with raw file paths

1. Download LoRA files from sources like Hugging Face or other repositories.
2. Upload LoRA to your environment (or point directly to a local path if your workflow supports it).
3. Load the Default Workflow in ComfyUI to establish a baseline for integration.
4. Add the LoRA Node: In the workflow, insert Load LoRA between the diffusion model and the CLIP encoder.
5. Configure LoRA Settings: Set lora_name to reference your LoRA file and adjust strength_model and strength_clip to control influence on the diffusion model and CLIP, respectively.
6. Connect the LoRA Node: Position the LoRA node between diffusion and CLIP nodes and wire inputs/outputs accordingly.
7. Test Your Workflow: Generate an image to verify the LoRA effect; tweak strengths as needed and optionally add trigger words to prompts for finer control.

Practical notes for different workflows

• SD1.5, SDXL, and Flux all support LoRA integration; follow similar steps, adjusting for each environment’s node layout and special considerations.
• For Flux, you may need to reload Flux-specific nodes before adding LoRA to Flux workflows.
• LoRA stacking is possible: you can combine multiple LoRAs using the LoRA Stacker to achieve nuanced effects, with independent strength controls for each LoRA.

How to test and fine-tune

Run the workflow and generate images to confirm that the LoRA’s style, character, or concept is being rendered as intended. If results aren’t satisfactory, adjust strength_model and strength_clip; you can also modify the prompt with trigger words used by initial LoRA authors to sharpen outputs.

Using LoRA in MimicPC with ComfyUI

LoRA integration through MimicPC streamlines downloading, uploading, and configuring LoRA models for SD1.5, SDXL, and Flux. The workflow typically includes loading the default setup, inserting the LoRA node, connecting it correctly, and testing the results, with optional stacking for advanced outcomes.

Images and videos to aid the workflow

Visual schematic showing how LoRA connects between the diffusion model and CLIP, with paths for lora_path and lora_name illustrated.

Short clip demonstrating inserting the LoRA node between diffusion and CLIP nodes in the ComfyUI workspace, including wiring inputs and outputs.

Table: key parameters and defaults

If the LoRA model file is not compatible or corrupted, check the file integrity and ensure compatibility with your ComfyUI version and the specific LoRA release you are using.

tags: #how #to #load #lora #in #comfyui