What Building an AI Outfit Tool Taught Me About Designing for Decisions
Summary: sarah wilson discusses the development of an AI tool for changing clothes in photos, highlighting key insights about user needs and product design. They emphasize the importance of reducing user input friction by leveraging existing images rather than text descriptions. Progressive control is recommended over maximum control, allowing users to start with basic visualization and refine as needed. The discussion also points out the need to clarify the limitations of AI-generated images, differentiating between visual exploration and physical fit. Finally, success is described not by the generated output, but by how effectively it aids user decision-making.
When we first thought about an AI tool for changing clothes in photos, the obvious product goal seemed simple: make the outfit change look convincing.
But that turned out to be only part of the problem.
A technically impressive image doesn't automatically help someone decide anything. Users don't necessarily need another beautiful fashion image. Often, they are trying to answer a much smaller question:
Would I actually explore this style further?
That distinction changed how I think about the product.
The User Isn't Starting With a Blank Canvas
A lot of generative image products begin with an empty prompt box.
For outfit exploration, that can create unnecessary work.
Imagine someone finds a jacket online that they like. They don't want to spend five minutes explaining its collar, length, material, buttons, color, and silhouette to an AI system.
They already have the reference.
And they already have another important piece of information: a photo of themselves.
That led us toward a much simpler interaction model:
Person photo + clothing reference → outfit visualization
I've been developing this idea through AI Outfit Changer, where the workflow starts with those two visual inputs rather than asking users to reconstruct an outfit entirely through text.
Prompts can still be useful, but they become a refinement layer rather than the starting point.
That sounds like a small product decision. In practice, it changes the experience quite a bit.
Reduce the Work Before the Result
One lesson I've taken from building AI products is that generation quality gets most of the attention, while input friction is easy to overlook.
But users experience both.
If someone has to learn prompt syntax, describe every visual detail, regenerate repeatedly, and figure out why a reference wasn't followed, a powerful model can still feel difficult to use.
For a focused tool, reducing that setup work can sometimes matter more than adding another control.
The question becomes:
What information does the user already have that we shouldn't make them describe again?
In this case, images carry much of that information.
A clothing reference communicates visual details that would be tedious to express in words. A person photo establishes the subject, pose, proportions, and scene.
The interface can then concentrate on the gap between them.
More Control Isn't Always Better
Another temptation when building an AI product is to keep adding settings.
Once you can control one thing, it's easy to imagine controls for ten more.
Some users absolutely want that depth. But every new setting also creates another decision before the user gets a result.
For outfit visualization, I've found it more useful to think in layers.
The first layer should answer the basic question quickly:
What could this outfit direction look like?
Only after that should users need to think about smaller refinements such as layering, accessories, materials, colors, or styling instructions.
This is a product lesson I'm starting to appreciate beyond fashion tools:
Progressive control is often more useful than maximum control.
Give people enough to begin. Expose complexity when they actually need it.
The Product Also Needs to Say What It Cannot Tell You
This may be the less exciting part of building an AI product, but I think it's one of the most important.
An AI-generated outfit image can be useful for visual exploration without being an accurate prediction of physical fit.
Those are different promises.
A generated image shouldn't be treated as confirmation that a particular size will fit, that fabric will behave exactly as shown, or that a color will look identical in the physical product.
Small garment details can change during generation too.
It would be easy to describe all of this as “virtual try-on” and let users assume more precision than the product can provide.
I'd rather make the boundary clear.
The useful job is narrower:
Help someone explore a clothing direction visually before making the next decision.
That may be less dramatic than claiming to replace a fitting room, but it's also a much more defensible product promise.
A Failed Generation Can Still Create Value
This was another interesting realization.
Suppose someone generates a preview of a long coat and immediately dislikes the silhouette.
Was that generation a failure?
From an image-generation perspective, maybe the user didn't get something they wanted to keep.
From a decision-making perspective, however, the tool worked.
The user learned something.
They can stop exploring that direction and move on to another one.
This changes the way I think about success metrics for generative products.
The best outcome isn't always:
“The user loved the generated asset.”
Sometimes it's:
“The user reached a decision faster.”
For tools used during brainstorming, shopping, design exploration, or early-stage planning, that distinction matters.
Don't Confuse Output With Outcome
Generative AI makes output extremely cheap.
We can create another image, another variation, another prompt, another version.
That makes it tempting to measure a product by how much it can generate.
But users rarely arrive because they need 50 outputs.
They usually arrive with a problem.
For an outfit tool, that problem might be uncertainty about a style. For another startup, it could be uncertainty about a landing-page layout, a product concept, a marketing direction, or a prototype.
The AI output is only useful if it reduces that uncertainty.
That's the broader product lesson I've taken from working on this:
Don't design the workflow around what the model can generate. Design it around the decision the user is trying to make.
Once that decision is clear, it becomes easier to decide which inputs matter, which controls deserve space, what limitations need to be communicated, and which features probably don't need to exist yet.
For us, the interesting question is no longer simply whether AI can put different clothes on a photo.
It's whether that visualization helps someone move from “I wonder” to “I know what I want to explore next.”
For a focused AI product, that can be enough.
.png?_a=BAVMn6E80){kind=small}