LoRAs — short for Low-Rank Adaptation — are one of the most useful tools in the Stable Diffusion ecosystem. They are small model add-ons that steer a base checkpoint towards a specific style, character, or subject without requiring you to fine-tune or replace the entire model. In ComfyUI, adding a LoRA to your workflow takes only a couple of nodes, but understanding how to use them correctly will save you a great deal of trial and error. This guide covers everything from downloading your first LoRA to stacking multiple ones in a single generation.
What LoRAs Are and Why They Matter
A full Stable Diffusion checkpoint can be several gigabytes in size. Training a new one from scratch requires significant computing resources. LoRAs solve this by training only a small set of weight modifications — typically between 10 MB and 200 MB — that sit on top of a compatible base model. The result is a highly specialised adjustment: a LoRA might teach a model to draw in the style of a particular artist, generate a specific fictional character consistently, or produce images in a photorealistic portrait style.
Because they are small and stackable, LoRAs let you mix capabilities without juggling multiple large checkpoints. You can layer a style LoRA with a detail-enhancement LoRA on a single base model and achieve results neither could produce alone.
Where to Download LoRAs
The most popular source for LoRAs is Civitai (civitai.com). It hosts thousands of community-created LoRAs covering styles, characters, clothing, lighting effects, and much more. Each listing shows which base model the LoRA was trained on — this compatibility information is critical, which we will cover shortly.
Hugging Face is another source, particularly for more technically focused LoRAs. Always check the model card for the recommended base checkpoint and any trigger words before downloading.
Where to Put LoRA Files
Once downloaded, place your LoRA files (which will typically have a .safetensors extension) in the following folder inside your ComfyUI installation:
ComfyUI/models/loras/Create the folder if it does not exist. ComfyUI scans this directory on startup. If you add a LoRA while ComfyUI is already running, click the refresh button in the model selector or restart the server to make it appear.
Adding a LoRALoader Node to Your Workflow
In ComfyUI, LoRAs are applied via the Load LoRA node (also referred to as LoRALoader). To add one to your workflow:
- Right-click on an empty area of the canvas and select Add Node.
- Navigate to loaders and select Load LoRA.
The Load LoRA node has two inputs: model and clip. It also has two outputs: MODEL and CLIP. In a standard workflow, your Load Checkpoint node provides both MODEL and CLIP outputs. You connect those outputs into the Load LoRA node, then connect the Load LoRA outputs to wherever MODEL and CLIP were previously connected — typically the KSampler and CLIP Text Encode nodes respectively.
The node also has a dropdown to select your LoRA file and two numerical fields: strength_model and strength_clip.
LoRA Strength Settings
The strength_model slider controls how strongly the LoRA influences the denoising model. The strength_clip slider controls how strongly it influences text encoding. A value of 1.0 applies the LoRA at full strength; 0.0 disables it entirely.
In practice, most LoRAs work well between 0.5 and 0.9 for model strength. Pushing above 1.0 often produces over-saturated or distorted results. A good starting point is 0.75 for both values, then adjust based on the output. Some style LoRAs are quite aggressive and need to be turned down to 0.4 or 0.5 to blend naturally with your prompt.
Trigger Words
Many LoRAs require specific trigger words in your positive prompt to activate their effect. These are words the LoRA was trained to associate with its style or subject. Without the trigger word, the LoRA may have little visible effect even at full strength. Check the Civitai listing for the recommended trigger words — they are usually listed prominently on the model page.
For example, a LoRA trained on a particular art style might require the phrase artwork in the style of XYZ somewhere in your prompt. Add it early in the prompt for the strongest effect.
Stacking Multiple LoRAs
ComfyUI makes stacking LoRAs straightforward. Simply chain multiple Load LoRA nodes in series: the MODEL and CLIP outputs of the first feed into the inputs of the second, and so on. The final node in the chain connects to the KSampler and text encoders.
When stacking, reduce individual LoRA strengths to avoid artefacts. If you are using two LoRAs, try setting each to around 0.5–0.6 rather than 0.8+. Three or more LoRAs at high strength will typically produce incoherent results.
SDXL vs SD 1.5 LoRA Compatibility
This is the most common source of confusion when starting out. LoRAs are not interchangeable between base models. A LoRA trained on SD 1.5 will not work correctly with an SDXL checkpoint, and vice versa. Attempting to use an incompatible LoRA will usually result in garbled output or a ComfyUI error.
When downloading a LoRA from Civitai, look for the Base Model tag on the listing. Common values are SD 1.5, SDXL 1.0, Pony (which is SDXL-based), and Flux. Match this to the checkpoint you are using. Flux-based LoRAs use a different loader node — the standard Load LoRA node works, but Flux requires its own workflow structure.
Common Mistakes
- Wrong base model: The most frequent issue. Always verify compatibility before downloading.
- Missing trigger words: The LoRA seems to do nothing. Check the model page for required prompt phrases.
- Strength too high: Values above 1.0 often cause distortion or colour artefacts. Stay within 0.5–1.0.
- LoRA not appearing in the dropdown: The file may be in the wrong folder, or ComfyUI needs to be refreshed. Ensure the file is in
models/loras/and not a subfolder within it unless ComfyUI Manager’s settings allow recursive scanning. - Corrupted download: A partial download may look like a valid file but cause errors on load. Re-download and verify the file size matches the listing.
For a full index of every ComfyUI guide on Serverman, see the ComfyUI complete guide and hub.
