Trying to design a Steampunk Zelda from the Twilight Princess game

DISCLAIMER: I understand that these image AI tools are being trained with artists' imagery, without their consent, and that is a whole legal and ethical can of worms I am not looking to open in this blog post. I think it is wrong and that we need legislation to protect artists' work. The purpose of this article is to simply share my personal experience with using generative AI. 

As someone who has a passion for cosplay and costume design, I was intrigued when I first heard about Midjourney AI, a generative AI technology that produces images based on keywords and themes. A lot of cosplayers love creating unique, anachronistic, or cross universe, versions of our favorite characters, and I wanted to see if AI could help me do just that by designing a steampunk version of Zelda, from the Twilight Princess game.  Below, I want to share my experience with Midjourney AI, and my thoughts on the AI technology.

Artificial intelligence, or AI, refers to the ability of machines to perform tasks that would normally require human intelligence, such as understanding natural language, recognizing patterns, and making decisions. AI has been around for decades, but has become more prevalent in recent years, as advances in computer hardware and software have made it possible to build more sophisticated AI systems.

Recently, tools like ChatGPT, Midjourney AI, DALL-E, and many others have become main stream and are already revolutionizing the way we work, live, and create. These tools are here to stay, so I think we should all figure out how to live with them and what their strengths and limitations are.

I first came across AI being used in costume design by Firefly Path. On her Instagram, she was sharing how she was trying to get Midjourney AI to create a dress design in her style that she could then create in real life. I thought this blending between the digital and physical world was so cool that I wanted to try it out for myself! (Read more about her project here)

For this experiment, I wanted to use Midjourney AI to generate costume inspirations for a steampunk Princess Zelda. The process was simple: I entered a few keywords related to the theme, and the AI generated four ideas. I could then either upscale one, use one as a prompt to get four more images,  have regenerate the images using the same prompt, or ask for a new prompt all together. I started out simple and direct, and what the AI produced was also pretty direct (though I wouldn't call the design simple):

Prompt: a full body image of steampunk princess zelda twilight princess cosplay, full-length

Not bad, but not quite what I was looking for. I then proceeded to go down a Midjourney rabbit hole and spent a few hours fine tuning the prompts, upscaling images, to see how far I could push the software and what kind of variety I could get out of it. Below is small sample of some of what I was able to achieve:

Prompt: a full body image of a steampunk princess zelda princess dress with lights, gold, pink, purple, and white in color, full-length, “Full-length portrait”, “Wide field of view”, neutral background, 8k, hd

Prompt: a full body image of steampunk princess zelda twilight princess, full-length, “Full-length portrait”, “Wide field of view”, neutral background

Prompt: a full body image of a steampunk princess zelda princess dress, holding a bow and arrow, gold, pink, purple, and white in color, full-length, “Full-length portrait”, “Wide field of view”, neutral background, 8k, hd

Prompt: a full body image of a steampunk princess zelda princess dress, holding a sword gold, pink, purple, and white in color, full-length,“Full-length portrait”, “Wide field of view”, neutral background, 8k, hd

Prompt: full body image of a steampunk princess zelda princess dress, holding a sword, with triforce, gold, pink, purple, and white in color, ultra quality, character design, full-length, “Full-length portrait”, “Wide field of view”, neutral background, 8k, hd

Prompt: beautiful steampunk princess zelda painting

Prompt: full body beautiful steampunk princess zelda painting, full-length, , “Full-length portrait”, “Wide field of view”, ultra quality, 8k, hd

Prompt: full body beautiful steampunk princess zelda painting, full-length, “Full-length portrait”, “Wide field of view”, ultra quality, 8k, hd — q 2 — v 4

Prompt: full body beautiful steampunk princess zelda painting, full-length, “Full-length portrait”, “Wide field of view”, ultra quality

I noticed that since I wasn't specific about Princess Zelda from Twilight Princess, it was sometimes generating an image with lots of blue and gold: very Breath of the Wild inspired. Midjourney AI allows you to use an external image as a prompt, so I was curious what would happen if I fed it one of the original concept arts from the game:

It definitely captured the painted style of the original art, which I liked, but it also ended up producing a lot of images that were too similar to each other in terms of design and style, which I guess would be expected:

Prompt: [image] + steampunk zelda painting ultra quality, 8k, hd full-length

Prompt: [image] + full body steampunk zelda beautiful painting ultra quality, 8k, hd full-length, “Full-length portrait”, “Wide field of view”, ultra quality, 8k

Prompt: [image] + full body steampunk zelda with circlet beautiful painting ultra quality, 8k, hd full-length, “Full-length portrait”, “Wide field of view”, ultra quality, 8k

All of these images were very different and could be used to make very different cosplays. Do I want something that looks like a Medieval knight? Or a Victorian lady? Or steamship captain? Or something more traditionally steampunk? I then started looking at some of favorite images in more details and trying to understand how I could bring one of these to life. And this is where I understood exactly why AI won't be replacing real artists anytime soon.

None, not one of the generated images, was actually feasible to create as is. They just didn't make sense. For example, look at these um, swords:

If I'm out adventuring, I want to have a nice utility belt with me with pouches, and potions, and other useful tools, but this doesn't make sense:

Or what about the lights on this skirt. They definitely look cool, but why are they there in the first place? What is their purpose? They seem to big and heavy to just be decorative gems so they should be some kind of battery/power source, but then why would they be on flimsy skirt?

And below is a sample of some other weird issues that I saw:

Instead of elf ears we have this weird looking hat/headdress situation

Not sure what's going on with the shoulder armor

What also really got to me is the fact that the triforce symbol was missing from pretty much every single image. It is an iconic symbol in all the Zelda games, and is always featured on Zelda's outfit, but here it was nowhere to be found. 

Basically, every single image was missing a story and context. None of the created characters had usable weapons, armors, tools, or looked like they were full, complete creations. If a human artist was tasked with creating such a character, I'm confident they would have understood how to make a character that made sense by creating a story, a world around them, and making them instantly recognizable. 

This is where I understood that as of the writing of this article, AI Image Generating tools have their limitations.  For example, Midjourney AI seemed to struggle with understanding the context of the fantasy theme, and some of the costume ideas it generated were simply unrealistic. Additionally, while the AI was good at generating a large volume of ideas quickly, it couldn't replace the human intuition and creativity that goes into designing a truly unique and meaningful costume.

However, I found that Midjoureny was incredible at creating images that can be used for inspiration. I believe that AI technology has the potential to help designers overcome creative blocks, generating new ideas, and assisting in the design process. In fact there are several images that I would like to combine together, and fine tune to potentially create an actually feasible cosplay. The tool was able create some very cool looking, and incredibly different designs for a character. I was definitely inspired and felt like I had some great ideas I could use as a starting point to create something really awesome.

2022 was the year that AI technology went mainstream and became easily accessible to everyone, not just engineers. It's controversial, it's fast, it's terrifying, but whether we like it or not, it's here to stay and only going to get smarter from here on out. It's a brave new world, and we have to learn to live in it.

A detailed walkthrough of how I constructed my Keqing costume, from the game Genshin Impact.

I chose to cosplay Keqing because I’m a huge fan of Genshin Impact, and I just think she has a gorgeous character design that I thought would be both fun and challenging to bring to life. This cosplay pushed me to try multiple new skills beyond just sewing, and I’m very proud of the final result. Below you will find a detailed overview of how I created this cosplay.

Capture from the Genshin Fashion Archive: https://gamesfashionarchive.net/

The costume is made up of the following 4 layers (all sewn by me):

I started by taking a basic princess seam block pattern, and slowly chipping away pieces until I had the main underdress pattern. The skirt and underskirt are simple circle skirt patterns which I made longer in back.  

Once that was done, I draped the overdress over the pattern and tweaked the placement of everything.  The front panels started out as a circle skirt that I then slowly patterned out into individual panels.  

All the muslin patterns were scanned and converted into SVGs in Illustrator so that I could then work  on creating the embroidery motifs and make sure each embroidery motif would be scaled and positioned correctly.

Before I could being sewing, I had to find right the right shade of purple silk that, when the dye was removed, would leave the right shade of pink. In the end, there was only 1 winner: Amaranth Purple Dupioni Silk from NY Designer Fabrics

Purple silk fabric was carefully dipped in dye remover to create the ombre effect of the final dress

Once all the SVGs were created and scale confirmed, each SVG was manually digitized into an embroidery pattern, as well as converted into a corresponding cut file for my Silhouette Cameo. 

It was challenging to scale everything for both devices so that when the cut file was placed onto the embroidery,  it would easily become an applique.

Each piece of embroidery had to be meticulously positioned over the ombre fabric, and both panels had to match and be symmetrical.

I drew out the shape of each panel, the embroidery design, as well as where I should

position the hoop. I then made sure that my embroidery was digitized to perfectly fit the space,

For the appliques that go over the embroidery, I wanted to make them from the same silk fabric as the dress. I used the invisible applique technique, and carefully stitched each applique by hand using invisible stitches.

STEPS TO MAKE: 

1. Cut out each motif from fusible interfacing 

2. Iron the interfacing to one piece of fabric 

3. Sew around all visible edges and fold inside out 

5. Carefully tack the finished applique by hand to the right side of the fabric over the embroidery

Once all the embroidery was finished and each pattern/panel piece was ready, the construction of the dress was the most straightforward part.   

The corset is a simple single layer coutil corset with German plastic boning steam shaped into position.

The underdress pieces are flatlined with silk organza, the bodice is lined with cotton, and there is some German plastic boning along several seams. The skirt is a circle skirt gathered with pleats and attached to the bodice.  

There were some interesting details on the front bodice pieces, so the bodice ended up being lined with  the lighter silk dupioni. The whole underdress also has a hidden boned layer, both in the front and the back, which allows it to easily keep its shape and provide a strong foundation for all the accessories.

The sleeves are made from 4 hand painted silk charmeuse panels. The fabric was first soaked in an antifusant, which turns the silk into a canvas and prevents paint from spreading. Then the design was carefully painted on each panel individually.  

The upper part of the sleeve consists of a piece of EVA foam wrapped in silk. It was attached to the painted panels by hand and then embellished with an embroidered panel, hand painted 3d printed pieces, ruffles, and cord, and of course elastic to keep everything on my arms.

All the accessories were made from a combination of EVA foam, 3D printed pieces, embroidery, and resin casting. The 3D 

Any accessories that needed to be flexible where made from EVA foam. They were cut, sanded, shaped, primed, and finally painted.

Each 3D print had to be scaled, printed, cleaned, thoroughly sanded, primed, painted and sealed. The bead dangles also had to have a brass eye pins attached to them so they could be turned into jewelry. The vision has LEDs glued to the back.

I used 3D printed models created by The Dangerous Ladies on an Anycubic Photon M3 Printer using ABS like Resin. Spray painted with Rustoleum

The front feather was a 6-part embroidery done on polyester organza with water soluble stabilizer. Aligning the hoop after each piece was done was the most challenging part.

I used Robison Anton thread on a water soluble stabilizer, stitched out on a Viking Topaz 40 sewing machine

The shoes started out as simple beige shoes which I dyed purple using Angelus leather dye.

The side motifs were made from flocked heat transfer vinyl that I then painted with Jacquard gold paint

The side panels were sewn from velvet, and the back heel piece was made from EVA foam.

Custom molds made from 3D prints using OOMOO 30-1A:1B silicone mold and Resin Obsessions Resin

Come read about my latest creations and other fun things!

My Experience with Midjourney AI

Keqing Cosplay Construction Details