Interview on AI Image Generation (Text-to-Image) Technology and Baidu Search Applications

Since 2023, AIGC technology has sparked a new wave of artificial intelligence, with AI painting becoming one of the most prominent applications of large models. These systems can generate images of various styles from user prompts, providing powerful tools for artists, designers, and creators.

In a recent InfoQ "Geek Talk" program, Baidu Search chief architect Tianbao was invited to discuss image generation technology. The conversation covered Baidu Search’s application scenarios, technical considerations, and practical deployment experience of text‑to‑image models.

2022 is regarded as the “year of text‑to‑image”. It saw the rise of open‑source models such as Stable Diffusion and closed‑source models like Midjourney, Adobe Firefly, and DALL‑E 3. Early versions struggled with quality (e.g., distorted faces), but later releases achieved much higher realism and style diversity.

For Chinese language models, preparing and cleaning Chinese‑semantic corpora is essential. Building high‑value image‑text pairs and removing low‑quality samples are necessary capabilities for effective image‑text alignment.

Baidu leverages its massive web‑wide Chinese corpus to collect diverse, high‑quality data. It applies relevance‑modeling algorithms (e.g., CLIPScore) to filter out mismatched pairs and supports over a thousand predefined visual styles, allowing users to specify both content and style in prompts.

The interview highlighted prompt engineering challenges: users must describe detailed content and desired style, while the model must handle long, attribute‑rich Chinese prompts. Strategies such as prompt expansion, rewriting, and controlling the order of style keywords were discussed to improve controllability and reduce bias.

Evaluation combines human feedback (RLHF), user interaction signals (clicks, likes, downloads) and machine metrics to assess image quality, aesthetic standards, and content consistency. Automated metrics help pre‑filter results, but final aesthetic judgments often require human review.

Looking ahead, the speakers expect rapid progress in video generation. While spatial quality is already strong, ensuring temporal consistency adds a new dimension of difficulty. They anticipate significant breakthroughs within the next one to two years.

At the end of the session, Baidu announced recruitment for visual algorithm engineers to join the text‑to‑image research team.