• Kiki and Mozart
  • Posts
  • Which Style Reference Version Should You Use In Midjourney?

Which Style Reference Version Should You Use In Midjourney?

PLUS: latest from Adobe, OpenAI, and Stability AI

In this newsletter, read about:

  • 🕵️‍♀️ Style Reference Versions in Midjourney

  • 🗞 News and Top Reads

  • 📌 AI Art Tutorial: Tips for Consistent Characters

  • 🎨 Featured Artist: Marianne

  • 🖼 AI-Assisted Artwork of the Week

  • 🤓 A Comprehensive Midjourney Guide

🕵️‍♀️ Style Reference Versions in Midjourney

Midjourney has now introduced FOUR different versions of their recently launched Style Referencing feature. Here’s how they differ according to the Midjourney announcement:

  • --sv 1 uses the original style reference algorithm from the first release of this feature.

  • --sv 2 uses the algorithm released two weeks ago and has been a default until recently.

  • --sv 3 is a new version of --sv 1 that is supposed to perform better, while also being more ‘vibey’ than --sv 2.

  • --sv 4 is a new version of --sv 2 with improved performance. It is also a default version as of today.

For me, the differences seem quite subtle, and I believe the preferences for these versions will be highly subjective and dependent on the specific use case. Today, I am going to compare these four versions based on two use cases: (1) a very clear pop art style and (2) photorealistic photoshoot images.

Pop Art style

As a style reference for the first experiment, I’m going to use the following image:

Prompt: An image of an attractive young woman standing next to a motorbike. She is dressed in the 1960s style. The image is in the pop art style. --ar 4:5 --stylize 300

Now, let’s see how different Style Referencing versions will process the following prompt.

Prompt: A man and a woman are standing next to a vintage car. --ar 4:5 --stylize 300 --sref https://cdn.midjourney.com/11560f72-f23c-49b1-9bf7-df35768ebe93/0_1.png

I will be adding --sv 1, --sv 2, --sv 3, and --sv 4 depending on the version I want to test.

--sv 1

--sv 2

--sv 3

--sv 4

In this specific case:

  • The results from --sv 1 are my favorite, but at least two images have errors in terms of positioning people next to the car.

  • In --sv 2 results, all images look more or less “correct”, but for some reason, in 3 out of 4 images, we see people from their backs. Anyway, image #4 looks pretty cool.

  • --sv 3 results include way too many errors and inconsistencies.

  • Similarly, in --sv 4, all but image #2, have some significant errors.

So, it looks like, in this example, all four versions performed quite well in style transferring, but the earlier versions (--sv 1 and --sv 2) were better at following the prompt and keeping the output image logically correct.

Now, let’s check how these versions will handle photorealism.

Fashion Photoshoot

The reference image for this experiment will be as follows:

Prompt: A minimalistic fashion photoshoot of a black woman. She is wearing a very bright green dress and orange shoes. The woman is sitting on the chair. The photoshoot is against a minimalistic white background. --ar 16:9 --stylize 400

Now, we’ll use this style reference image with the following prompt:

Prompt: A fashion photoshoot of a Mexican woman in a bright orange dress. --ar 16:9 --sref https://cdn.midjourney.com/2cf9e86b-e0ea-464b-965f-396d2b26e6d3/0_0.png --stylize 400

--sv 1

--sv 2

--sv 3

--sv 4

I think in this example, you can more apparently see how --sv 3 performs similarly to --sv 1, while --sv 4 performs similarly to --sv 2:

  • Both --sv 1 and --sv 3 rely more on a new prompt and less on the reference image. Specifically, both of these versions take dress color from the prompt and don’t mix it with the dress color from the reference image. On the other hand, --sv 3 seems to mix a little bit with the women’s appearances, making a Mexican woman look more like a Black woman.

  • Both --sv 2 and --sv 4 mix the dress colors. Moreover, --sv 2 seems to ignore the minimalist white background in the reference image, adding a lot of bright colors to the background. At the same time, I like the generations from these two versions more as they appear more creative and attention-grabbing.

To Wrap Up

These two small experiments alone are not sufficient to draw any definitive conclusions about the performance of different Style Referencing versions. Additionally, we have yet to explore the impact of varying the amount of style guidance provided through the prompt.

However, even from these limited use cases, it's evident that the outcomes from different versions can vary significantly, with preferences strongly tied to the specific use case. Therefore, if you're finding it challenging to achieve the desired results with Style Referencing, it's certainly worthwhile to experiment with various versions, as well as adjusting prompts, the style weight (--sw), and the overall level of stylization (--s).

Happy Prompting!

🗞 News and Top Reads

  • Adobe announced a suite of new AI tools:

    • GenStudio emerges as a standout release, offering users the ability to generate content, manage brand assets, monitor campaign performance, and optimize workflows within a single application.

    • A novel generative AI assistant debuts, equipped to address technical inquiries, automate processes, and simulate business outcomes.

    • Additionally, Adobe collaborates with Microsoft, facilitating seamless integration of marketing insights and workflows from Adobe and Microsoft applications into Microsoft’s Copilot AI for Microsoft 365, aiming to enhance marketer efficiency.

  • OpenAI advances the development of Sora, its cutting-edge video-generation tool:

    • Valuable input from visual artists, designers, creative directors, and filmmakers who tested the tool has been gathered.

    • The company showcased numerous new examples, illustrating the diverse applications of Sora for creatives.

    • Additionally, OpenAI is venturing into the film industry by presenting its text-to-video tool to Hollywood.

  • Stability AI turned out to be not that stable:

    • First, key Stable Diffusion researchers have left the company.

    • Then, Emad Mostaque resigned from his role as CEO of Stability AI and from his position on the Board of Directors of the company to pursue decentralized AI.

📌 AI Art Tutorial: Tips for Consistent Characters

In this video tutrial, Nolan from Future Tech Pilot shares some important tips on creating consistent characters in Midjourney. Learn how to create the same character with different face expressions, what you can get from using color as character reference, and how to combine character reference and style reference.

Meet Marianne, aka @pinkwritinglife – she's not your average AI artist! Her style is all about making a statement. From thought-provoking to downright beautiful, her creations stick with you long after you've seen them. And guess what? Pink is her signature color, but there's nothing ordinary about it. Dive into her vibrant world over on Instagram @pinkwritinglife for more eye-catching art!

🖼 AI-Assisted Artwork of the Week

🤓 A Comprehensive Midjourney Guide

To get a link to a comprehensive Midjourney guide, please subscribe to this newsletter. The guide is a dynamic document, which I intend to keep up-to-date with the latest Midjourney updates.

Share Kiki and Mozart

If you enjoy this newsletter and know someone who might also appreciate it, please feel free to share it with them. Let's spread the word about AI art and introduce more people to this fascinating field!

Join the conversation

or to participate.