Designing the Experience Around Clad's Analysis Model
A model can predict a color season, but photo quality, result explanation, and recommendation logic determine whether someone can trust and use the result.
The central capability in Clad is personal color analysis from a photo. A correct season label, however, is not a complete product result. The user needs to submit a suitable image, understand why the analysis reached its conclusion, and apply the result to colors or products they can actually choose.
In many AI products, the experience before and after inference is treated as supporting work. In personal color analysis, where input conditions and interpretation have significant influence, those surrounding decisions determine whether the output feels credible.
Result quality starts at upload
Lighting, white balance, filters, shadows, and camera processing can make the same person look different across photos. A generic upload prompt gives the user no way to understand what the analysis needs. When the result is weak, neither the user nor the product can easily distinguish an input problem from a model problem.
The upload step should explain, in concrete terms:
- use neutral or natural light;
- avoid strong filters and makeup that changes skin tone;
- choose a front-facing image where the face is large enough to inspect;
- expect to retake the photo if the input does not meet the requirements.
Where practical, the product should validate brightness, face size, and obvious image quality before analysis. Asking for a better photo early is more useful than producing a sophisticated answer from unsuitable evidence.
The process should feel predictable
Behind the interface, Clad may prepare the image, run analysis, normalize the result, generate a palette, and map recommendations. Users do not need this architecture diagram, but they should know what is happening, how long it may take, and how to recover if something fails.
A staged progress state can reduce uncertainty better than one indefinite loading screen. Errors should distinguish between a photo that needs to be replaced and a temporary service problem that can be retried.
Explain the evidence, not only the label
A label such as “Warm Spring” is memorable but insufficient. The result should describe the observed color tendencies, the useful range of value and saturation, and the kinds of contrast that may be less suitable.
A confidence number alone does not solve the explanation problem. The interface needs to clarify what the number means, why confidence may be lower, and whether a different input could improve it. Showing how uncertainty is handled is more credible than presenting false precision.
Keep palettes and recommendations on one logic
The palette is the fastest way to make an abstract analysis tangible. Grouping colors by use, presenting them at a comparable size, and separating clothing from makeup contexts can help users understand the result without reading a long report.
Beauty recommendations should follow the same color logic. A list of generally popular products makes the analysis and recommendation feel disconnected. The product should be able to explain why a shade was selected and where it sits within the recommended palette.
Measure what happens after the result
Analysis completion does not show whether the result was useful. More informative signals include whether users read the reasoning, explore or save the palette, continue into recommendations, or rerun the analysis. Qualitative feedback should ask which explanation created confidence and where interpretation became difficult—not only whether the season label felt correct.
The product challenge in Clad is not to display a stronger model. It is to manage imperfect inputs, explain a probabilistic result, and carry that result into a coherent set of real choices.