High-Efficiency Diffusion Models for On-Device Image Generation and Editing with Hung Bui - The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Recap

Podcast: The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence)

Published: 2025-10-28

Duration: 52 min

Summary

Hung Bui discusses his transition from leading AI research in the Bay Area to establishing a world-class AI lab in Vietnam, focusing on high-efficiency diffusion models for mobile devices. He emphasizes the importance of adapting models to fit the constraints of mobile hardware while achieving superior performance.

What Happened

In this episode, host Sam Charrington welcomes Hung Bui, the new VP of technology at Qualcomm following their acquisition of VinAI Research, known for its impressive research output in AI. Hung shares his journey into the field of artificial intelligence, which began nearly 30 years ago during his PhD studies on multi-agent systems. He reflects on his early fascination with AI and the Turing test, a curiosity that propelled him through notable positions at Google DeepMind and Adobe Research, before eventually returning to Vietnam to set up the first AI research lab in the country.

Hung describes the challenges he faced in building a research team in Vietnam, particularly the need for talent. He explains how he managed to attract experienced researchers from abroad while also nurturing young local talent through an AI residency program, which was the first of its kind in Southeast Asia. This dual approach allowed VinAI to leverage the knowledge of seasoned professionals while fostering a new generation of AI researchers, significantly contributing to the lab's reputation for efficiency in deploying AI models on mobile devices.

A key highlight of the discussion is Hung's emphasis on the development of high-efficiency diffusion models. He shares insights into their efforts to create a seven billion parameter model, which faced criticism for being too large for local GPUs. In response, they successfully created a model with fewer than four billion parameters that outperformed the larger variant. This achievement underscores Hung’s commitment to making advanced AI tools accessible on mobile platforms, reflecting a broader trend in the industry towards optimizing models for practical use in real-world applications.

Key Insights

• Hung Bui's career trajectory highlights the evolution of AI from academic curiosity to impactful industry applications.
• VinAI Research successfully combines experienced researchers and young talent to foster innovation in AI.
• Reducing model size can lead to better performance, as demonstrated by Hung's team's experience with diffusion models.
• The push for on-device AI emphasizes the need for efficiency and accessibility in generative AI technologies.

Key Questions Answered

What motivated Hung Bui to return to Vietnam to establish an AI lab?

Hung Bui explains that returning to Vietnam was a significant opportunity for him to make an impact in his home country. He describes a common aspiration among professionals working abroad to contribute to their native countries. However, he also acknowledges the risks involved in setting up a world-class AI lab in a place like Vietnam, where the environment for AI research was still developing.

How did Hung Bui address the talent challenge in Vietnam's AI sector?

To tackle the talent challenge, Hung Bui implemented a strategy that involved hiring experienced researchers willing to relocate to Vietnam, alongside nurturing local young talent. He initiated the first AI residency program in Southeast Asia to attract bright young minds, allowing them to gain practical experience while contributing to the lab's research efforts.

What are diffusion models, and why are they significant in AI?

Diffusion models are a type of generative model used for tasks such as image generation and editing. Hung's team focused on optimizing these models for mobile devices, which is crucial for making advanced AI technology more accessible. Their work demonstrated that a smaller model could outperform a much larger one, highlighting the importance of efficiency in AI applications.

What insights did Hung Bui share about the development of AI in Vietnam?

Hung emphasized that developing AI in Vietnam involved significant challenges, particularly regarding investment and talent acquisition. He shared his experience of transitioning from the Bay Area's robust AI ecosystem to Vietnam, noting both the opportunities and risks. Ultimately, he believes that the lab's focus on generative models and efficiency has positioned it well within the global AI landscape.

How does VinAI's approach to AI research differ from traditional models?

VinAI's approach emphasizes high efficiency and adaptability, particularly for mobile applications. The lab's research direction, influenced by Hung's previous work, focuses on deep generative models, which are essential for modern AI advancements. This focus allows them to create models that not only perform well but are also optimized for the constraints of mobile hardware.