Style Blending: "Create" new styles by mixing style LoRAs

CivitAI is packed with thousands of style LoRAs from specific artists or artistic movements. Beyond simply replicating a single style, you can mix several of them to “create” something fresh and new.

I’m writing this article to raise awareness about the creative freedom style LoRAs offer in image generation; I want to show that they are much more flexible than people think and aren’t just for imitation.

First, let’s introduce the LoRAs we’ll be using. I’ve picked a few at random that aren't very similar to each other to see what we can achieve:

Test 1: All Strengths at 1.0

The first step is to see what happens if we use all models at a strength (weight) of 1.0.

On its own, it looks okay, but there’s a problem when you run everything at 1.0: the LoRAs start "fighting" each other. At the same time, they overlap on certain traits, which ends up "breaking" some concepts. In this image, for example, it tended to add nudity and background details were lost.

If we look only at the style, we have something good on our hands (a “new” style), but we still have conceptual issues. While generating these, several images came out wrong, with elements appearing that weren't even mentioned in the prompt, mostly due to the bias of some of the LoRAs used.

Test 2: All Strengths at 0.8

If we drop everything to a strength of 0.8, things start to stabilize a bit more while still maintaining a solid style. We could keep lowering the weights until we find the sweet spot between all models.

We can choose to set all LoRAs to the same strength, but that isn't always the best approach. Ideally, you want to go for a “a little bit of this, and a lot of that” strategy.

There’s a rule that can be useful (though it depends on how aggressive the LoRA is) which says that the sum of the weights shouldn't exceed 1.0. For example, if you use 3 LoRAs, the strengths would be 0.3 + 0.3 + 0.4 = 1.0.

The goal here is to stay close to the base structure of the checkpoint you're using to prevent the image from breaking. However, with current architectures like SDXL (Illustrious), this rule takes a backseat because they have more parameters and much more flexible limits.

Personally, I don't follow this rule since I haven't had issues exceeding it, but I’m mentioning it because someone might find it useful.

Experimenting with Variations

Next, we’ll test different weights between the same LoRAs to find different combinations: (All images were made with the same parameters and seed)

Swapping a LoRA

We’ve obtained some very interesting variations using the same 5 LoRAs, but there comes a point where you stop seeing noticeable changes. So, why not swap one LoRA for something completely different?

Note: In the first image, we can see artifacts on the pointed ears due to the "Haunted Mourning" LoRA. This is one of the errors you can run into when combining LoRAs; that specific model is quite strong and seems to be breaking the "elf ears" concept. Simply lowering the weight even further should fix it, as 0.8 seems to be excessive here.

I could spend all day testing different models and weights, but this clearly demonstrates the ability to create fresh styles using existing style LoRAs.

UPDATE 01/11/2026

I forgot to mention something important!!
Some checkpoints make the LoRA effects almost non-existent, while others let their full potential shine. Use a neutral and flexible checkpoint.
An example of a checkpoint I wouldn't use is WAI Illustrious.
An example of a checkpoint I would use is Plant Milk - Walnut.

Conclusion

To be honest, it’s sometimes tricky to find that perfect balance where you love the style and don't run into technical issues (artifacts). Of course, you can use fewer models; using 3 LoRAs instead of 5 greatly reduces the chance of error, and maybe that LoRA you had at 0.2 wasn't even that necessary.

For me, this is the true potential of artist LoRAs: creating something different using real foundations. I like to see it as taking existing work as a reference to draw something “new,” rather than just imitating it.