How AI-Powered Image Generation Revamped Baidu App’s Avatar Customization

Introduction

As AI technology develops, many products try to combine AI for feature upgrades. Designers face the problem of how to integrate AI with scenario needs to provide a smooth AI‑native design experience.

This article uses the AI avatar project as an example, introducing how we analyzed user pain points, leveraged internal capabilities, found AI innovation breakthroughs, and explored a design‑driven innovation path.

Part 1: Why innovate the avatar AI

1. Baidu App avatar business overview

Avatar is a basic capability of Baidu App, supporting users to set avatars, skins, homepage backgrounds, outfits, etc., aiming at personalization. Two main problems exist:

Function lacks attraction: The workflow (select → preview → apply) is single and lacks fun.

Avatar resources iterate slowly: Updates depend heavily on designers; manual drawing is costly and time‑consuming.

2. Avatar user analysis

Research and data analysis revealed two findings:

Personalization needs are hard to meet: Free avatar resources cannot satisfy users’ personalized demands.

Younger users love AI: The user base is young, interested in intelligent, fun products and has higher demand for customization.

3. Understanding internal capability support

Baidu WENXIN AI image‑generation is mature, supporting free generation of multiple images from descriptive prompts.

4. Design opportunity

We asked: Can avatar integrate WENXIN AI image generation so users can generate images themselves for avatars?

We analyzed feasibility and value from user, business, and design perspectives:

User side: Enrich avatar play, improve fun and attraction.

Business side: Attract more users, increase avatar user volume and penetration.

Design side: Explore real AI application scenarios, accumulate AI‑native design experience, and reduce designer cost.

Part 2: How to achieve AI avatar innovation breakthrough

We iterated three phases: Phase 1 rapid validation, Phase 2 innovative breakthrough, Phase 3 data‑driven optimization.

Step 1: Rapid validation

We surveyed domestic and overseas AI‑generated products. Competitors’ creation pages are similar, with high input barriers.

Goal: low‑cost, fast launch to test user interest.

AI avatar flow: click entry → start creation → generate image → preview avatar.

Design challenges:

Simplify creation process.

Reduce step loss.

Design challenge 1: Simplify creation process

Solution: let users pick keyword combinations to generate images, simplifying the workflow.

Design challenge 2: Reduce step loss

Solution: shorten the path from image generation to avatar preview.

We chose the shorter‑path solution.

Phase 1 results

Daily avatar settings increased nearly threefold; AI avatar conversion funnel was high, ranking second among avatar types.

Step 2: Innovative breakthrough

Compared to Phase 1’s semi‑automatic generation, Phase 2 aimed to let users input their own prompts for higher personalization, but the input barrier was higher.

We identified two design difficulties: (1) how to lower the barrier for high‑quality image generation, and (2) how to restructure the avatar flow.

For (1) we explored:

Provide high‑quality prompt examples.

Assist users with prompt refinement, advanced prompt library, formula‑guided word combination.

We also explored a double‑layer framework separating guidance content and input panel, allowing users to scroll guidance like a feed while keeping the input stable.

Multiple design proposals were created and tested; most users preferred proposal 1.

Final solution combined image examples with prompt generation, removed explicit style selection, and used intelligent style recognition.

Step 3: Continuous iteration

Phase 3 focused on balancing image preview and step reduction. We introduced motion effects to connect pages and reduce jumps.

Two directions were examined:

Short primary path, secondary functions after core action.

Preview image only, then post‑action.

Three detailed proposals were created; user testing favored proposal 1.

Final flow: during generation, show re‑generate, save, etc.; after generation, auto‑scroll image and launch avatar preview panel, shortening the path.

Data after optimization showed a significant increase in daily successful avatar conversions.

Conclusion

This case study demonstrates how designers can combine AI capabilities with user‑centered design to create smooth AI‑native experiences, using a three‑step process of rapid validation, innovative breakthrough, and data‑driven refinement.

Key takeaways:

Avoid assumptions; always view problems from the user’s perspective.

User testing is essential to uncover hidden pain points and generate actionable design insights.

Keep design solutions simple and intuitive, reducing cognitive load and conversion steps.