A LoRA is a small model file that modifies a larger image generation model. Instead of downloading a whole new checkpoint to get a particular art style, character or concept, you load a LoRA on top of the checkpoint you already have and control how strongly it applies. They're one of the most popular ways to customise local AI image generation, and ComfyUI makes them easy to use once you know where the files go.
This guide covers installing LoRAs, the Load LoRA node's settings, using more than one LoRA at once and fixing the common problems. Forge PC hasn't tested image output or performance for this article; the steps come from ComfyUI's official documentation. If you haven't set up ComfyUI yet, start with how to install ComfyUI on Windows.
What you need
A working ComfyUI install with at least one checkpoint model.
A LoRA that matches your checkpoint's model family. A LoRA is trained for a particular base model. Civitai's training documentation, for example, says its SDXL training produces a LoRA usable with any SDXL checkpoint, and its SD1.5 training produces one usable with any SD1.5 checkpoint. An SDXL LoRA won't be the right match for an SD1.5 checkpoint, so check the base model listed on the LoRA's download page.
Safetensors files where possible. Hugging Face describes safetensors as a simple format for storing tensors safely, as opposed to pickle, while still being fast.
If your graphics card is short on memory, see how to run ComfyUI on a low-VRAM GPU.
Step 1: Put the LoRA in the right folder
Download the LoRA file.
Copy it into ComfyUI/models/loras. ComfyUI's documentation says the Load LoRA node detects models in this folder, including subfolders, so you can organise LoRAs into folders by style or model family.
If ComfyUI was open while you copied the file, ComfyUI's models guide says to restart the app or refresh so the lists update.
ComfyUI's docs say all LoRA variants, including LyCORIS, LoHa, LoKr and LoCon, are used the same way.
Already have LoRAs in another app's folder?
If you've used another Stable Diffusion app, you don't need to copy everything. ComfyUI can read extra model folders configured in extra_model_paths.yaml in the ComfyUI folder. ComfyUI provides an extra_model_paths.yaml.example file to rename and edit. For the ComfyUI desktop app on Windows, the docs give the config file location as C:\Users\YourUsername\AppData\Roaming\ComfyUI\extra_models_config.yaml. Restart ComfyUI after saving the change.
Step 2: Add the Load LoRA node
ComfyUI's official LoRA tutorial uses the dreamshaper_8 checkpoint with the blindbox_V1Mix LoRA, and you can load its example workflow by dragging the example image into ComfyUI. To add a LoRA to your own workflow:
Start with a basic text-to-image workflow that has a Load Checkpoint node.
Double-click an empty part of the canvas, search for Load LoRA and add it.
Connect the MODEL output of Load Checkpoint to the model input of Load LoRA.
Connect the CLIP output of Load Checkpoint to the clip input of Load LoRA.
Connect Load LoRA's MODEL output to where the checkpoint's model used to go (usually the KSampler).
Connect Load LoRA's CLIP output to your positive and negative CLIP Text Encode prompt nodes.
Pick your file in the lora_name drop-down.
Click Queue, or press Ctrl + Enter (Cmd + Enter on a Mac), to generate.
The key point is that both the model and the CLIP connections must pass through the LoRA node. If the KSampler is still wired directly to the checkpoint, the LoRA won't be used.
Step 3: Set the LoRA strength
The Load LoRA node has two strength settings:
strength_model sets how strongly the LoRA changes the model's weights. ComfyUI says higher values make the LoRA's style stronger.
strength_clip sets how strongly it changes the CLIP text embeddings, which affect how your prompt is interpreted.
ComfyUI's node documentation gives a range of -100 to 100, but says values between 0 and 1 are typical for everyday image generation. Start with the value recommended on the LoRA's download page if there is one, then lower it if the effect is too strong or distorts the image. Some LoRAs also list trigger words to include in your prompt, so check the download page.
Step 4: Use more than one LoRA
You can combine LoRAs, for example a style LoRA with a character LoRA. ComfyUI says the Load LoRA node supports chain connections, so you link several in series:
Add a second Load LoRA node.
Connect the first LoRA node's MODEL and CLIP outputs to the second node's model and clip inputs.
Connect the second node's outputs to the KSampler and prompt nodes.
Adjust strength_model on each node to balance their influence.
ComfyUI's multiple-LoRA tutorial chains the blindbox_V1Mix and MoXinV1 LoRAs this way to blend their styles. When combining several, lowering each one's strength is a good starting point, so no single LoRA overwhelms the image.
Model-only LoRAs
ComfyUI also has a LoraLoaderModelOnly node, which loads a LoRA without needing a CLIP model and applies it to the model alone. ComfyUI says it reads from the loras folder and from any extra paths in extra_model_paths.yaml. If a workflow you download uses it, only the model line passes through the LoRA.
Troubleshooting
The LoRA doesn't appear in the list. Check it's in ComfyUI/models/loras (or a configured extra path), then restart or refresh ComfyUI.
No visible effect. Make sure the KSampler takes its model from the Load LoRA node, not the checkpoint, and check you've included any trigger words.
Strange or broken images. The LoRA may be for a different base model than your checkpoint, or the strength may be too high.
If you'd rather use a simpler app, we compare ComfyUI with AMD's Amuse in ComfyUI vs Amuse, and our best local AI image generator guide covers the alternatives. For hardware advice, see the best PC for Stable Diffusion and FLUX.
Our verdict
Using a LoRA in ComfyUI takes three things: put the file in ComfyUI/models/loras (subfolders are fine), add a Load LoRA node between your Load Checkpoint node and both the KSampler and prompt nodes, and set strength_model and strength_clip, usually between 0 and 1. Chain several Load LoRA nodes to combine styles, and always match the LoRA to your checkpoint's model family. Use extra_model_paths.yaml if you want to share LoRAs with another app. Forge PC hasn't tested image output or performance for this guide.