How to Use LoRAs in ComfyUI: Load LoRA, Stacking, Strength and Fixes (2026)

A LoRA is the quickest way to teach a big image model something new: a character style, a product look, an art style or a camera effect. The file is small, often just tens to a few hundred megabytes, and you can swap it in and out without touching the base model.

ComfyUI makes LoRAs easy once you know where the files go and how the Load LoRA node is wired.

This guide shows how to use LoRAs in ComfyUI step by step, how to stack several of them, the strength settings that actually matter, model-only LoRAs for Flux, Wan and other newer models, and fixes for every common error.

What a LoRA Does

LoRA stands for Low-Rank Adaptation. Instead of retraining the whole model, a LoRA stores a small set of extra weights that nudge the base model towards something specific.

When ComfyUI loads a LoRA, it patches those weights into the model (and usually the text encoder) for the current run. The base checkpoint file on disk never changes, so you can use one checkpoint with dozens of different LoRAs.

Common uses include:

• Styles: watercolour, pixel art, 3D toy, film photo, anime styles.
• Characters and mascots: your own made-up character kept consistent across images.
• Concepts and objects: a product, a vehicle design, a clothing style.
• Helpers: detail enhancers, lighting LoRAs, and speed LoRAs that cut the number of steps (Lightning, Turbo and distillation LoRAs).
• Control and editing: newer models ship LoRAs for pose, depth or edge control, like the MOSS-Image control LoRAs.

LoRA Types and Which Node to Use

ComfyUI’s own LoRA tutorial confirms that LoRA variants such as LoCon, LoHa and LoKr can be loaded with the regular Load LoRA node.

Step 1: Download a LoRA That Matches Your Model

1. Check your base model. Open your workflow and note the checkpoint, for example an SD 1.5 model like DreamShaper 8 or an SDXL model.
2. Search a model site. CivitAI and Hugging Face are the two big LoRA libraries. Filter by base model (SD 1.5, SDXL, Flux.1 D, Wan and so on).
3. Read the model page. Note the trigger words, the recommended strength and any sampler or CFG advice.
4. Prefer .safetensors. Avoid old .ckpt or .pt files from unknown uploaders, since those formats can contain code.
5. Check the licence. Some LoRAs ban commercial use. That matters if you sell images or use them on a monetised site.

The official ComfyUI LoRA example uses the DreamShaper 8 checkpoint (SD 1.5) with the blindbox_V1Mix LoRA, which turns prompts into cute 3D toy figures. It is a good first test because the effect is obvious.

Step 2: Put It in the Right Folder

After copying the file, press R in ComfyUI (refresh) or restart it. The file should now appear in the lora_name dropdown. Subfolders are fine; they show up as folder/file.safetensors.

New to ComfyUI? Our ComfyUI install guide covers Desktop, Portable and manual installs, and the ComfyUI beginner’s guide explains every core folder.

Step 3: Build the Load LoRA Workflow

Start from the default text-to-image workflow (Workflow > Browse Templates, or the default graph). You only add one node.

1. Add the node. Double-click the canvas, search Load LoRA and add it.
2. Connect MODEL. Drag the MODEL output of Load Checkpoint into the model input of Load LoRA.
3. Connect CLIP. Drag the CLIP output of Load Checkpoint into the clip input of Load LoRA.
4. Rewire the sampler. Connect the MODEL output of Load LoRA to the model input of KSampler.
5. Rewire the prompts. Connect the CLIP output of Load LoRA to both CLIP Text Encode nodes (positive and negative).
6. Pick the file and strengths. Choose your LoRA in lora_name, set strength_model and strength_clip to 1.0 for a first test.
7. Run. Add the trigger word to your prompt and press Queue (Ctrl+Enter).

The order is always: Load Checkpoint → Load LoRA → (KSampler + CLIP Text Encode). VAE still comes straight from the checkpoint (or your separate VAE loader).

Step 4: Trigger Words and Prompting

Most LoRAs are trained with one or more trigger words. Without them, the effect is weak or missing.

• Put the trigger word first or near the start of the positive prompt.
• Keep the rest of the prompt simple on your first try, so you can see what the LoRA adds.
• Do not type <lora:name:1>. That is Automatic1111 syntax. Core ComfyUI ignores it; the node does the loading.
• Style LoRAs often work without a trigger word, but adding the style name still helps.

Example prompt for the blindbox test: chibi, full body, a girl in a yellow raincoat holding an umbrella, standing on a street, cute 3d toy, blind box style, soft studio lighting.

Strength Settings Explained

Load LoRA has two sliders. They do different jobs.

A good habit: run the same seed at 0.4, 0.7 and 1.0 and keep the lowest value that still gives you the look you want. Lower strength keeps more of the base model’s quality.

How to Use Multiple LoRAs

You can stack LoRAs by chaining Load LoRA nodes. The ComfyUI docs show exactly this pattern.

1. Add a second Load LoRA node.
2. Chain it. Connect the MODEL and CLIP outputs of the first Load LoRA into the model and clip inputs of the second.
3. Connect the last node only. The final Load LoRA’s MODEL goes to KSampler and its CLIP goes to both text encoders.
4. Lower the strengths. With two LoRAs try 0.6 to 0.8 each; with three, 0.4 to 0.6 each.
5. Add both trigger words to the positive prompt.

Stacking a style LoRA with a character LoRA works well. Stacking two strong style LoRAs often fights, giving muddy images. If that happens, lower the weaker one or drop it.

Model-Only LoRAs (Flux, Wan, Qwen-Image)

Newer models use large separate text encoders (T5, UMT5, Qwen and others). Most LoRAs for these models only change the image or video model, so you use LoraLoaderModelOnly.

1. Load the diffusion model with Load Diffusion Model (or a GGUF loader, see our GGUF in ComfyUI guide).
2. Add LoraLoaderModelOnly and connect the MODEL output into it.
3. Send its MODEL output to the sampler (KSampler, or the model input of your guider / ModelSamplingFlux / ModelSamplingSD3 node).
4. Leave CLIP alone. The text encoder loader connects to CLIP Text Encode as before.
5. Chain more LoraLoaderModelOnly nodes the same way if you need several.

Speed LoRAs (Lightning, Turbo, distillation) are usually model-only too. They change the sampler settings you need: fewer steps and CFG around 1. Always follow the model card. Our Qwen-Image-Edit Lightning guide and P-H3-Image 2-step LoRA guide show real examples.

Organising a Big LoRA Collection

• Use subfolders by base model: loras/sd15/, loras/sdxl/, loras/flux/, loras/wan/. It stops you picking the wrong family.
• Rename files to include the trigger word, for example sdxl_watercolor_trigger-wtrclr.safetensors.
• Keep a notes file next to each LoRA with the source URL, strength and licence.
• Share folders with Forge or A1111 using extra_model_paths.yaml instead of copying gigabytes twice.
• Delete what you never use. Hundreds of files slow the dropdown and waste disk space.

Want to make your own? See how to train a LoRA with the Kohya GUI, or train inside ComfyUI with ZenTrainer.

Troubleshooting

Still stuck after an update? See how to update ComfyUI safely, since old builds sometimes miss LoRA support for new models.

My Take

LoRAs are the best value upgrade in ComfyUI. Learn the one-node pattern first: Load Checkpoint, Load LoRA, then the sampler and prompts. After that, stacking and model-only LoRAs are small steps.

Keep strengths modest, always add trigger words, and organise files by base model. That avoids nine out of ten “my LoRA doesn’t work” problems.

FAQ

Where do LoRA files go in ComfyUI?

In ComfyUI/models/loras/. Press R or restart ComfyUI so they appear in the Load LoRA node.

What strength should I use for a LoRA?

Start at 0.7 to 1.0 for both strength_model and strength_clip, or the value on the LoRA’s page. Lower it if the image looks overdone.

How many LoRAs can I use at once?

There is no hard limit. Two or three is common. Lower each strength as you add more, or they will fight each other.

Do I need trigger words in ComfyUI?

Usually yes. Most character and concept LoRAs need their trigger word in the positive prompt. Style LoRAs often work without one.

What is the difference between Load LoRA and LoraLoaderModelOnly?

Load LoRA patches both the model and the CLIP text encoder. LoraLoaderModelOnly patches only the model, which is what Flux, Wan and most newer model LoRAs need.

Can I use SDXL LoRAs with SD 1.5 or Flux?

No. A LoRA only works with the model family it was trained on. Mixing families gives warnings and little or no effect.

Do LyCORIS, LoHa and LoKr files need a special node?

No. ComfyUI’s regular Load LoRA node loads them.

Can I use LoRAs with video models?

Yes. Wan and other video models have their own LoRAs, loaded with LoraLoaderModelOnly. See our ComfyUI image to video guide.

Sources

• ComfyUI docs: LoRA tutorial (Load LoRA, models/loras, DreamShaper 8 + blindbox example)
• ComfyUI docs: Applying multiple LoRAs
• ComfyUI docs: LoraLoader node reference
• ComfyUI on GitHub
• LoRA: Low-Rank Adaptation of Large Language Models (paper)