Building AI for Creators | Luma & Phota Labs

The podcast explores the evolving relationship between artists and AI creative tools through a conversation with Matt Tancic from Luma and Zach Xia from PhotoLabs. Both experts discuss how creativity is fundamentally about storytelling and directing tools—whether traditional or AI-powered—rather than mastering them. Matt reflects on how his views have evolved since discussing NERF three years ago; while he initially framed AI as a tool for artists to execute their vision, he now sees AI's role as even more central across multiple modalities like video and images. Zach emphasizes that technology should help artists express creativity rather than forcing them to spend time mastering complex workflows. The conversation addresses the significant shift in photography from capturing the decisive moment to post-capture editing, where generative AI has enabled photographers to enjoy more creative freedom in post-production while remaining authentic to the original moment. Both speakers highlight the challenge of transitioning from research to product-building. Researchers naturally push toward advancing technology and solving novel problems—like rendering esoteric fonts in generated text—while creators need practical solutions to immediate problems like removing backgrounds or adjusting lighting. The speakers note that successful product design requires balancing technological advancement with user-centric problem-solving. A key insight emerges around personalization and subjective evaluation. Unlike traditional metrics in research, creative tools must account for highly subjective preferences. The speakers describe surprising user behaviors, such as people combining PhotoLabs with video models to preserve identity across video frames, or brands bringing their own guidelines to influence outputs. They emphasize that good evaluation requires both human judgment from tastemakers during development and post-launch user feedback to understand satisfaction. The transcript reveals tensions in how to represent creative work and assets. Rather than settling on a single universal representation, the speakers argue that the right representation depends on what kind of control and modification is needed in the future. A poster with changeable text needs different representation than pixel-level image modification or 3D scene navigation. Both the model and the human user must be able to understand and manipulate these representations. The discussion explores controllability as essential to professional creative work. Text-based prompts alone are insufficient; creators need visual controls like scribbles, region selection, and video-to-video transformations. Importantly, Zach introduces the idea that models should ask clarifying questions rather than working one-way, mirroring how professional studios interact with clients. The speakers describe significant latent demand for AI photography tools among people without access to professional equipment or skills. This represents a different user segment than professional photographers—people who want high-quality documentation of important life moments but lack the resources or expertise to achieve it traditionally. Interestingly, professional photographers and AI generation can coexist; photographers can use AI to explore creative ideas impossible in physical reality. On the future separation between elite and average artists, both speakers agree that while AI is raising the baseline quality for everyone, the gap between great artists and average users will likely widen. Great artists understand their creative intent deeply, can uniquely combine tools in unexpected ways, and understand how to direct agents effectively. Finally, the speakers address personalization and style, arguing that style cannot be easily defined through keywords or explicit description but must be learned from data—both from a user's existing work and from their reactions to generated outputs. Style also changes over time and varies by project, suggesting that true personalization requires understanding both broader user preferences and task-specific needs.