Why does the image break down when I use multiple LoRAs?

Applying multiple LoRAs at once can cause their characteristics to interfere with each other. Start by lowering each LoRA's strength below its standalone value, and if the image still breaks down, remove them one at a time to isolate which LoRA is causing the problem.

What should I check when a LoRA doesn't seem to have any effect?

First confirm the file is in the correct folder (models/loras/ for ComfyUI, models/Lora/ for AUTOMATIC1111) and that you restarted or refreshed the UI. Then check whether you've included the trigger word in your prompt and whether the strength is set too low.

What strength (weight) should I set for a LoRA?

0.6-0.8 is the standard range. If the effect is too weak, raise it to 0.9-1.0; if the image breaks down, lower it to 0.4-0.6. When combining multiple LoRAs, lower each one's strength below its standalone value and adjust while checking the output.

How to Use LoRA: Installation, Trigger Words, Strength, and Combining Multiple LoRAs

You’ve got a LoRA file, but now what — where does it go, how do you write the prompt to activate it, and why does the image fall apart when you stack a few together? These are common stumbling blocks.

This article walks through installing and actually using LoRA in practice. For an explanation of how LoRA works under the hood, see What Is LoRA? — if you just want to know how to use it, keep reading here.

Where to Place LoRA Files

LoRA files (extension .safetensors) go in the designated folder of whichever UI you’re using.

ComfyUI

ComfyUI/models/loras/

Place the LoRA file here, then restart ComfyUI or refresh the model list in the browser to have it recognized.

AUTOMATIC1111

stable-diffusion-webui/models/Lora/

Same idea — after placing the file, restart the UI or refresh the model list (the 🔄 button) to pick it up.

You can organize files into subfolders (e.g. models/loras/character/, models/loras/style/), and the folder structure carries over directly into the UI’s LoRA list.

How to Call LoRA in Your Prompt

AUTOMATIC1111: the `<lora:name:weight>` syntax

In AUTOMATIC1111, you call a LoRA by writing this in your prompt:

<lora:filename:strength>

For example, to use anime_style.safetensors at strength 0.8, use the filename without its extension:

masterpiece, best quality, 1girl, 32yo woman, <lora:anime_style:0.8>

Selecting the LoRA from the tab below the prompt box inserts this syntax automatically, which avoids filename typos.

ComfyUI: the LoRA Loader node

ComfyUI doesn’t use inline prompt syntax — instead you add a LoRA Loader node to the workflow.

CheckpointLoader
  ↓ MODEL, CLIP
LoraLoader (select LoRA file + strength_model + strength_clip)
  ↓ MODEL              ↓ CLIP
  KSampler ← CLIPTextEncode (prompt)

The LoraLoader node has two strength parameters: strength_model (effect on image generation) and strength_clip (effect on prompt interpretation). A simple starting point is to set both to the same value, then adjust them independently only if you notice a difference in behavior.

Finding and Setting Trigger Words

Some LoRAs require a specific word (a trigger word) in the prompt to activate — or to activate fully.

How to find it

Check the distribution page: Civitai model pages often have a “Trigger Words” field listing the word to copy
Look at sample prompts: sample images on the distribution page frequently include the trigger word in their listed prompt
Check the model card (Hugging Face): the README-style model card sometimes shows example captions used during training

If the trigger word is unclear, try generating without one first, and if the effect feels weak, go back and re-check the distribution page.

How to set it

Add the trigger word like any other prompt term, comma-separated:

masterpiece, best quality, 1girl, 32yo woman, sks style, <lora:example_style:0.8>

If adding the trigger word makes no noticeable difference, the LoRA may have been trained in a way that doesn’t require one (for example, using caption dropout during training).

Strength (Weight) Guidelines

Strength is generally specified in the 0-1 range (values outside that range are technically possible, but 0-1 is usually sufficient).

Strength	Effect
0.3-0.5	Subtle application. Mostly preserves base model output while slightly adding LoRA characteristics
0.6-0.8	Standard application. Recommended range for most LoRAs
0.9-1.0	Strong application. LoRA characteristics stand out, but image quality may degrade

Start around 0.7 and adjust from there — raise it if the effect is too weak, lower it if the image breaks down.

Combining Multiple LoRAs

Using several LoRAs at once is possible, but it’s where most problems show up.

Common symptoms

Composition or texture breaks down
One LoRA’s characteristics dominate excessively, or another’s don’t show up at all
Details like faces and hands lose consistency

Fix 1: Lower each LoRA’s strength

Stacking multiple LoRAs while keeping each at a high, standalone-level strength (such as 0.8) tends to make them interfere with each other and break the image. When combining LoRAs, start each one lower than you would use it alone (roughly 0.3-0.6) and adjust one at a time while checking the output.

Example: Style LoRA (0.6) + Character LoRA (0.3)

The workable range depends on how the specific LoRAs interact, so find it by inspecting the output rather than by any numeric total.

Fix 2: Remove them one at a time to isolate the cause

To find which LoRA is causing the breakdown:

Remove one LoRA and regenerate — check whether the issue is gone
If not, put it back and remove a different LoRA, then check again
Once the issue disappears, the LoRA you just removed is likely the cause

Once you’ve identified the culprit, lower its strength further or try swapping in a different version of the LoRA.

Fix 3: Consider node order (ComfyUI)

In ComfyUI, chaining multiple LoraLoader nodes can produce different results depending on the connection order. If you’re not getting the effect you expect, try swapping the node order and comparing.

Base Model Compatibility

LoRAs are trained against a specific base model architecture, so they don’t work correctly across different model families.

LoRA trained on	Works with	Doesn’t work with
SD1.5 family	SD1.5 family	SDXL family, other families
SDXL family	SDXL family	SD1.5 family, other families
Other families (e.g. Flux)	Same family	Different families

Symptoms of a mismatch:

Images break down significantly, or look completely unchanged from a normal generation
An error appears when loading
The LoRA has no discernible effect at all

Always check the “supported models” note on the distribution page and make sure it matches the base model you’re using.

When Your Local GPU Runs Out of VRAM

Combining multiple LoRAs or generating at high resolution increases VRAM usage, and a local GPU can run short. When VRAM runs out, generation may fail partway through or slow down dramatically.

If your local GPU isn’t enough, renting a cloud GPU by the hour is another option — useful when you want to experiment with high resolutions and multiple LoRAs without worrying about VRAM limits.

PR RunPod クラウドGPUでAI画像生成 RunPodを始める →

Summary

Place LoRA files in models/loras/ for ComfyUI, or models/Lora/ for AUTOMATIC1111
AUTOMATIC1111 uses <lora:name:strength> syntax; ComfyUI uses a LoRA Loader node
Find trigger words on the distribution page’s description or sample prompts
Start strength around 0.6-0.8 and adjust based on effect and artifacts
When combining LoRAs, lower each LoRA’s strength, and remove them one at a time to isolate issues
Always check base model family compatibility (SD1.5 / SDXL / other) between the LoRA and your checkpoint

For how LoRA works under the hood, see What Is LoRA?. To train your own LoRA, see the LoRA Training Guide.