AWS re:Invent 2025 - The path to Multimodal Artificial General Intelligence (STP102)

Summary of AWS re:Invent 2025 - The Path to Multimodal Artificial General Intelligence (STP102)

Introduction to Multimodal AGI

• Current language models like GPT are incredibly capable, but lack understanding of the physical world
• To interact with the real world, AGI needs to understand visual and spatial intelligence, not just linguistic intelligence
• Luma AI's mission is to build multimodal AGI that can generate, understand, and operate within the physical world

Key Features of Ray 3 Model

• Ray 3 is Luma's state-of-the-art video generation model that can:
  • Reason about and refine its own output in real-time to match instructions
  • Export high-quality HDR and EXR files for VFX and post-production
• Key strengths of Ray 3:
  • Realistic physics simulations, fluid dynamics, and explosions
  • Coherent world motion, depth, and scale
  • Anatomically accurate character animation and movement
  • Detailed lighting effects, textures, and surfaces

New Capabilities of Ray 3

• HDR and EXR output support for advanced color control and compositing
• Reasoning model that annotates and refines output to best match instructions
• Visual annotation mode for sketching and directing model output
• Draft mode for low-res previews to conserve credits

Ray 3 Core Features

• Image-to-video: Turning static images into cinematic motion
• Text-to-video: Bringing ideas to life through written instructions
• Modify video: Editing existing videos by changing start frames or prompts
• Key frame animation: Seamless transitions between start and end frames
• Video looping: Creating endless, seamless loop animations

Advanced Style Adherence

• Ray 3 preserves the look, feel, and aesthetic of the original input
• Maintains consistent lighting, camera effects, color palette, and surface details
• Able to translate a variety of styles, from moody intensity to retro kung fu

Business Impact and Use Cases

• Content demand has outpaced human production capabilities
• Generative video unlocks speed and scale for enterprises:
  • Consumer packaged goods and e-commerce: Generating product videos, lifestyle shots, and personalized assets
  • Film and episodic production: Accelerating pre-production, storyboarding, and VFX workflows
  • Gaming: Enhancing 2D and 3D game asset creation and animation
  • Advertising: Producing high-quality, targeted ads at scale

Real-World Examples

• Coca-Cola 2025 holiday ad campaign created by a small team in 30 days
• 32-second animation produced for $630 vs. $95-180K with traditional methods
• Brands able to generate dozens of creative directions and iterate quickly

The Path to Multimodal AGI

• Luma's goal is to solve multimodal general intelligence, going beyond just video generation
• Capturing and simulating the physical world is key to empowering anyone to produce cinematic-quality content