PROMPT GUIDE

How to Combine Multiple LoRAs (Strength Balance and Fixing Mixed or Broken Pictures)

A single LoRA is already useful, but combining two or more, such as a character LoRA and a style LoRA, widens what you can express at a stroke. On the other hand, simply adding up the strengths can break the picture or mix styles into something different. This guide sums up how to write several LoRAs, how to balance their strengths, and how to fix a broken picture.

Last updated: 2026-10-22 approx. 3 min read

日本語版

The short answer

When you use several LoRAs, line up <lora:name:strength> entries in the prompt. Start each one weak, around 0.5 to 0.7, and do not stack too many of the same kind (style with style and so on). If the picture breaks, lower the strengths and remove them one by one to find the cause. Also check that the base model family (SD1.5 or SDXL) matches.

How do you write it when you use several LoRAs?

Line up as many <lora:name:strength> entries as you want to use in the prompt.

In AUTOMATIC1111-style screens, you write something like <lora:filename:0.6> inside the prompt. When you use several, you simply line them up one after another. On some screens, clicking a LoRA in the list inserts this form automatically. In ComfyUI, you connect LoRA loader nodes and set each strength. How and where to write it differs between tools, so check the explanation for the screen you use.

1girl, solo, school uniform, classroom, <lora:character-name:0.7> <lora:style-name:0.5>

1girl, solo, long hair, outdoors, <lora:style-name:0.6> <lora:detail-tweaker:0.4>

Rinka's tips

For LoRAs that need a "trigger word," also write it in the body of the prompt. Check the LoRA's description page
A file name that contains non-ASCII characters or symbols may fail to load. Choosing from the LoRA list is the surest

How do you balance the strengths?

Start each one around 0.5 to 0.7 and raise it little by little if that is not enough.

The higher you set a LoRA's strength, the more it takes effect, but if you strengthen several by adding them up, the picture moves further and further from the base model's picture and breaks. As a rule of thumb, do not let the total get too large, and lower LoRAs of the same kind as if they were sharing the strength. Keeping the role of each kind in the next table in mind makes the balance easier to decide.

Type of LoRA	Role	Rough strength (to start)
Character	Matches a specific character's face, hair and clothes	0.6 to 0.8
Style	Changes the art style: painting, lines, color mood	0.4 to 0.7
Detail and adjustment	Slightly tidies details and contrast	0.2 to 0.5
Pose, clothes, props	Adds only that element	0.5 to 0.8

Steps for combining two LoRAs

Choose LoRAs of the same model family (SD1.5 with SD1.5, SDXL with SDXL)
First use just one at around 0.6 and check that it works as expected
Add the second from around 0.4 and compare with the seed fixed
If it breaks, lower both. Raise only the one you want to come out stronger a little
Write down the strengths of the good combination and the name of the model you used

For the basic use of a single LoRA and how to think about strength, see How to Use LoRA (Syntax, Strength and Trigger Words) →

To compare several strength levels at once, see How to Use the XYZ Plot (Compare Settings Side by Side) →

Anime girl in a blazer smiling in a classroom, a base image before adding extra styles — 基準の絵

A soft watercolor-style version of the classroom girl, showing how a style can change the look — 基準の絵

Illustration usage rules

How do you fine-tune the strength?

Line up pictures with the same seed at several strength levels, and you will find the right spot.

The effect of a LoRA can change a lot with a change of just 0.1. Fix the seed of a picture you like, and compare by changing only one LoRA's strength to 0.3, 0.5 and 0.7 and lining them up. A comparison feature such as an XYZ plot lets you line them up at once. The trick is to leave the strengths of the other LoRAs alone and change only one at a time.

What you want to see	What to change	What to fix
How the character's face comes out	The strength of the character LoRA	The seed and the strength of the style LoRA
The strength of the style	The strength of the style LoRA	The seed and the strength of the character LoRA
How well the two work together	A picture without one and a picture with both	The seed, prompt and size

What should you do when styles mix or the face breaks?

Lower the strengths and remove them one at a time to find the LoRA that causes it.

Common failure	Cause	Fix
The art styles mix into something different	Two style LoRAs are stacked	Narrow the style to one, and remove the other or set it to 0.3 or lower
The face does not become the character LoRA's face	The style LoRA overwrites the face	Lower the style LoRA's strength and raise the character LoRA a little
Colors get dark or burned	The total strength is too large	Lower everything by 0.2 to 0.3
The whole picture breaks	You are mixing LoRAs of a different model family	Do not mix SD1.5 ones with SDXL ones. Check the supported family
One of the LoRAs has no effect at all	There is no trigger word, or the strength is too low	Add the words as the LoRA description says and raise the strength

For a review when the art style varies every time, see Why your AI art style keeps changing, and how to keep one consistent look →

For how to think about keeping the same character consistent, see How to Keep the Same Character in AI Art (Face, Hair and Outfit) →

FAQ

Frequently Asked Questions

How many LoRAs can I use at the same time?

There is no fixed limit, but the more you add, the easier it is for the picture to break. Try two first and add only the ones you need.

Does the order I write LoRAs in matter?

It is said to make little difference, but the total strength and the combination matter much more. When you want to compare orders, line them up with the same seed.

Can I mix SD1.5 and SDXL LoRAs?

No. A LoRA is used to match the model family it was made for. Check the supported family in the LoRA description.

Can I set the strength to 1.0 or more?

You can, but it breaks more easily. Use the recommended range in the author's description as a guide, and trying below 1.0 first is recommended.

I use the LoRA as distributed but the picture is different.

The model, VAE, sampler or Clip skip may differ. If a sample has generation info, match all the settings and try again.