AI Costumes: Even if You Can Generate Them, They Can Still Fail
A beautiful image and a functional costume are two different things.
The image was generated beautifully.
Glossy white and pale purple materials, red and gold decorative knots, and a waist cloth that drapes long on one side. It matched the character and background well, and when I first saw it, I thought, 'I've created a very unique costume.'
Even so, when I re-examined it as a costume, I couldn't call it a success.
The failure I'm referring to here doesn't mean the image was distorted or failed to generate. It looked attractive enough. The problem was that it hadn't reached the unique costume structure I was aiming for.
When I removed the character and background and broke down the costume alone, the cause became clear.
What looked unique was mainly the surface decoration. The basic structure supporting the costume had barely changed from a common swimsuit shape.
In this article, I will compare that failed example with a successful one created by changing my way of thinking.
It is a record of the journey to realizing the difference between 'having' generative AI draw clothes and 'having' it design a structure that makes the clothes functional.
Why is it a 'failure' even though it was generated?
At first glance, the initial image looked highly polished.
The color scheme was cohesive, the decorations were detailed, and even the footwear was unified. If viewed as an image piece, it wouldn't be strange to call it a success.
At first, I also thought that this 'beauty as a finished image' was synonymous with 'success in costume design.'
However, the more I looked back at it, the more an inexplicable sense of discomfort remained.
Where is this costume supported?
Which parts are necessary as a costume, and which parts are just decorations?
Is the long waist cloth on one side part of the structure? Or is it just a decoration added to an existing swimsuit?
Even if it looks gorgeous from the front, what does it look like from the side or back?
At the time, I thought the cause of the discomfort was 'a lack of decoration,' so I added more knots and chains. However, even by making it look more luxurious, the fundamental problem did not disappear.
This is because the problem was not in the amount of decoration, but in the basic structure of the costume.
Having a novel appearance is not the same as having a novel costume structure.
When I removed the character and background, the structure of the costume became visible.
When the character is wearing it, your eyes are drawn to the amount of information in the face, hair, pose, and background.
Therefore, instead of just releasing the original image of the character wearing it, I created a design board that organizes the costume by its individual components.
This task is not just image editing to erase the person. I read the top, bottom, waist cloth, decorations, and footwear from the wearing image and reconstructed each as a reference material for the costume alone.
This is to temporarily remove the impression created by the body and background and confirm, 'What role does this part play when viewed on its own?'
Even elements that worked as charms in the finished image become easier to distinguish as functional parts or visual decorations when replaced with a blueprint. When evaluating failed images, I have started to include this 'process of looking only at the costume'.
What I want you to pay attention to is not the luxury of the decorations, but the basic form of the top and bottom.
A failed example where only the costume is broken down, excluding the person and background. Although there are plenty of Japanese-style decorations, the main structure remained a standard bikini shape.
When viewed as a costume alone, both the charm and the problems become visible at the same time.
Red and gold flower knots, thin chains, fabric with a pale luster, and a long waist cloth. These have sufficient charm.
However, what remains after removing the decorations is a top that is easily understood as an existing swimsuit and a standard triangular bottom. The long red waist cloth was also not a part that replaced the function of the bottom, but a hanging cloth added to the decoration on one side.
In other words, even if the theme and decorations were unique, the main structure that makes the costume work remained conventional.
This is not about the 'bikini shape being bad'.
What I was aiming for was not a design that added Japanese-style decorations to an existing swimsuit, but a costume where the support method and component configuration themselves were different. The failure was that the structure did not change in line with that goal.
The problems could be organized into the following four points.
The existing bikini shape is the main structure
Japanese-style decorations do not change the structure
The long waist cloth is an additional decoration, not a functional part
Surface decoration takes precedence over support paths
By creating this board, I was finally able to put my discomfort into words.
Simply adding a theme to the decoration does not change the structure of the costume.
Where I was going wrong was thinking in terms of 'clothing names'.
At the time, I was coming up with ideas based on clothing category names, such as 'Japanese-style swimsuit,' 'luxurious swimsuit,' or 'unusual swimsuit'.
I would add materials, colors, decorations, and long fabrics to that category. The atmosphere would change, but the foundation of the design would remain the existing swimsuit shape that the generative AI easily understands.
At least in my production process, the more I started thinking of it as a 'swimsuit,' the more it would often end up returning to a swimsuit-like shape.
So, I increased the number of negative words, such as 'do not make it a normal swimsuit' or 'do not make it triangular'.
However, even if you increase the number of forbidden words, if it is not decided what is supported where and which part is responsible for what, it will not result in a stable structure.
It wasn't just the performance of the model that was the problem.
The entry point of my instructions was too focused on 'what kind of clothing name it is'.
So, I decided to categorize costumes not by name, but by role.
Parts that support the upper section
Parts responsible for the front
Parts responsible for the back
Anchors that serve as left and right fulcrums
Tension paths that transmit force to each part
Connections that prevent shifting or rotation
Light sheer materials that create movement
Decorations integrated into the structure
Think of these component parts for each function as "modules."
By dividing them into modules, you can limit the scope of any necessary corrections. If the front is too wide, you only fix the front. If the back is unstable, you review the support path for the back. You don't have to recreate the parts that are already working, like the top, along with them.
Even when prompting generative AI, specifying "maintain the left and right anchors and make only the front module thinner" makes it clearer what to keep and what to change, rather than using vague corrections like "make the entire costume more creative."
I didn't want to increase the number of difficult terms. I wanted to clarify which parts exist and for what purpose.
Instead of breaking down the names of the clothes, I broke down the roles that make the clothes work.
Designing not just the front, but the sides and back
The next thing I created was not a plan to add decorations to an existing swimsuit.
It is a structure where the top, front, and back are designed as separate functional parts, each supported by anchors and thin tension paths.
In the failed example, the basic shape of the bottom existed first, and the long loincloth was added to the side of it.
In the successful example, I designed the front and back as separate parts from the beginning.
There are aquamarine anchors on the left and right. The red hanging cloth on the front is supported by thin paths independent of the left and right anchors. The hanging cloth on the back also uses the same anchor system, but via a different path than the front.
The important thing is that the front and back are not connected as a single piece under the body.
From the front alone, you can only see the shape of the front.
From the back alone, you can only see the shape of the back.
Only by looking at the full side view can you confirm whether the front and back parts are independent or if they are connected in a hidden position.
From this experience, the side view became not just supplementary material, but an important diagram for verifying the structure.
Independent parts don't just need to be placed in mid-air.
Where is the fulcrum? Where does the force flow from and to? Which part stabilizes the front, and which path supports the back? It must be possible to read these things.
In the successful example, the thin lines extending from the left and right anchors explain the relationship between the parts.
I also gave the decorations a structural role.
The aquamarines on the left and right are not only color accents but also anchors that support the front and back. The central jewelry piece is also used not just to draw the eye, but as a connection point where multiple paths converge.
Instead of increasing the number of decorations, give them a role.
By doing so, I no longer had to treat decorations and structure separately.
In the successful example, the decorations do not hide the structure, but rather work to make the structure readable.
“Generation success” and “design success” were different things
The difference between the failed example and the successful example is not the amount of fabric or the lavishness of the decorations.
In the failed example, the existing shape came first, and the theme was added to the surface. In the successful example, I assembled it from the functional parts that make the costume work.
The long waist cloth in the failed example was an additional decoration. In the successful example, the front and back parts each have a clear role.
In the failed example, the support paths were buried within the decorations. In the successful example, the anchors and tension paths are readable.
In the failed example, the impression from the front was the focus of evaluation. In the successful example, I verified the front, full side, and back separately.
From this comparison, the meanings of “generation success” and “design success” diverged for me.
Generation success means that an image is output and it looks coherent as a visual.
Design success means that each part has a role, the support and connections can be explained, and the structure does not break down when viewed from multiple directions.
Both are important, but they are not the same thing.
Furthermore, there are stages to design success. Even if it works on a technical board, parts may spread out when put on a character, or thin support lines may be omitted when a background is added.
Therefore, I have stopped treating structure boards, three-view drawings, wearing images, and location images as the same level of passing, and have started checking them at each stage.
The successful examples shown in this article were not decided by a single front-facing image alone. I confirmed the separation of the front and back in a full side view, checked the width and connections from the back, and fed those results back into the costume-only board.
I prioritized maintaining the same structure under different conditions over creating a single beautiful image.
From this experience, three principles that can be applied to other costumes have remained.
Evaluate the beauty of the finished image and the viability of the costume separately
Look at not just the front, but also the full side and back views
Determine the parts, support points, and connection relationships before the costume name
AI costume failure is not just when it fails to generate. Because it is generated beautifully, structural failures are sometimes overlooked.
Turning failed images into the next blueprint
Failed images do not need to be deleted and forgotten.
What was attractive about it?
What was different from the goal?
Which parts functioned, and which parts were just decoration?
By removing the person and background and breaking down only the costume, the places that need to be fixed next become visible.
In this lab, we will record not only the finished images but also the failures that led to them.
We plan to cover methods for dividing costumes into functional modules, designing anchors and tension paths, verifying front, side, and back views, separating the roles of reference images, and how to create correction policies from failed images.
Do not have generative AI draw clothes, but have it design them from the structure.
Do not discard failed images; turn them into the next blueprint.
That is what I want to continue doing at 'Mizu | Generative AI Costume Design Lab'.
Next time: Why it reverts to a normal swimsuit
Next time, I will take up the topic of 'Why does AI revert creative costumes to normal swimsuits?'
Why do they return to existing forms even when themes or decorations are increased?
I plan to organize, using practical examples, how the "category name" at the beginning of a prompt fixes the form, in addition to the model's own quirks.
What is the biggest problem you face when creating outfits with AI?
If you could let me know in the comments, I will use it as a reference for future articles.