Where AI meets design judgment

Walking through the first version of image translation, I kept feeling like something was off. The translated text was clean, but it didn't belong — stripped of the visual personality that made the original content worth looking at.

I went back through hundreds of real posts and started mapping what I saw. Most text on rednote images isn't body copy — it's headlines, handwritten notes, stylized labels. Bold, delicate, outlined, colored. Each carrying a visual personality the translation had flattened into a single default style.

Working with my PM, I defined a new labeling framework around one question: does it preserve the original's visual intent? That meant defining visual dimensions the model had never been asked to recognize: Text scale, Stroke style, Color, Text background, Font weight, Script type, Layout density.

We translated those dimensions into concrete, checkable attributes. The matrix below shows four examples: Large title, Has stroke, Has color, and Has background. From there, we used the framework to label, test, and refine the model until its output began to match the visual intent we had defined.

With the labeling framework in place, the model's text-style recognition accuracy increased from 29.6% to 61.8%.

Still, one product question remained: should video covers also be included in the translation flow? Translating them could create a brief mismatch: the cover would appear in English, but text inside the video would still be Chinese. Because consistency across the feed mattered more, I chose to include video covers in the same translation flow, accepting the mismatch once users opened the video.

The A/B test evaluated the full translation experience, including video covers. In the English-language segment, feed engagement time increased by 4.3% and feed card CTR by 1%.

The same approach extended to voice dubbing. When we noticed regional accents carrying over across languages—for example, a Sichuan-accented Chinese speaker becoming a Sichuan-accented English voice—I treated it as a cross-cultural labeling question, not just a technical problem.

Instead of assuming the same accent should carry over unchanged, I worked with the ML team to adjust how much of it the dubbed voice should retain in another language. The question was not whether the model could reproduce the accent, but whether it would carry the same meaning for a different audience.

Defining "good" for an AI isn't always about pixels. Sometimes it's about what feels human.

When we shipped voice comments, transcripts appeared by default so users could read a comment before choosing whether to listen. That kept the experience quiet, but raised a harder question: if users could already read the comment, what would make the voice itself worth hearing?

The tag was our solution: one small label that gives you a reason to listen before you've heard anything.

My PM and I developed a classification framework covering timbre, dialect, emotion, music style, and ambient sound. A tag appeared only when one of those qualities was distinctive enough to call out.

The framework gave us a shared way to review early model outputs, decide which qualities deserved a tag, and iterate as the rollout expanded. After launch, a user called the "sweet voice" tag accurate and said it made them tap and keep listening.

The tag was the visible result. The larger takeaway was a new design responsibility.