I Got Early Access to Runway 4.5 + Kling AI Demo
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
In this episode of The Next Wave podcast, Matt Wolfe teams up with AI writer Maria Gharib to test the latest generation of AI video models, specifically diving deep into Runway Gen 4.5, Kling AI 2.6, and Kling O1. (00:30) The duo puts these cutting-edge tools through rigorous real-time testing, generating videos from quirky prompts like "monkey on roller skates" and "T-Rex wearing mittens making brownies on Mars" to evaluate their capabilities. (03:53) They explore the practical applications for businesses, test lip-syncing and audio generation features, and experiment with image-to-video conversion. The episode also addresses the controversial McDonald's AI-generated Christmas ad that sparked widespread criticism, discussing the broader implications of AI in corporate marketing. (42:17)
- Core focus: Hands-on testing and comparison of the newest AI video generation models, with practical insights for content creators and businesses looking to leverage these tools effectively
Speakers
Matt Wolfe
Matt is a prominent AI content creator and the host of Future Tools newsletter, with a popular YouTube channel where he regularly tests and reviews the latest AI technologies. He frequently receives early access to new AI models and tools, making him a trusted voice in the AI community for practical, hands-on insights into emerging technologies.
Maria Gharib
Maria is an AI writer and newsletter creator who specializes in exploring the practical applications of artificial intelligence tools. She brings a creative perspective to AI testing and is particularly interested in how these technologies can be applied to real-world business scenarios and content creation.
Key Takeaways
Detailed Prompts Drive Better Results
Both Matt and Maria discovered that generic prompts produce subpar results across all AI video models. (04:02) When they tested a basic prompt like "create an ad for a meal prepping business," the result showed a woman awkwardly eating food that appeared to be falling out of her mouth. However, when they used ChatGPT to generate detailed cinematographic prompts with specific camera angles, lighting, and scene descriptions, the quality improved dramatically. The key takeaway is that AI video models require rich, descriptive prompts that include technical details like camera movement, lighting conditions, and specific actions to produce professional-quality content.
Starting with Images Produces Superior Video Quality
Matt emphasized a crucial insight from his extensive testing: "I tend to find that when you're using video models, you almost always get a better result if you start from an image and generate a video off the image." (31:59) This approach leverages the existing visual composition, lighting, and detail of a well-crafted image, allowing the AI to focus on animation rather than generating everything from scratch. For businesses looking to create marketing content, this suggests a workflow of first creating high-quality images (using tools like Midjourney) and then animating them with video AI.
AI Video Should Complement, Not Replace Human Creativity
The McDonald's AI-generated Christmas ad controversy highlighted a critical principle for professional applications. (44:18) Matt advocates that "AI needs to be used as like an assistive tool, not as like the main tool." The most effective approach is using AI for 10-20% of a project—perhaps for shots that would be cost-prohibitive or impossible to film traditionally—while maintaining human actors, real VFX, and traditional production values for the majority of the content. This hybrid approach avoids the "AI slop" perception while leveraging AI's unique capabilities.
Model-Specific Strengths Require Strategic Selection
Through comparative testing, the hosts revealed that different AI video models excel in different areas. Kling 2.6 demonstrated superior physics and more cinematic movement, as evidenced in their dragon comparison, while Runway 4.5 showed better consistency in certain scenarios. (27:00) Kling also includes audio generation, while Runway focuses solely on video. For professionals, this means developing a strategic approach to tool selection based on project requirements rather than relying on a single platform.
Current Models Still Have Significant Limitations
Despite impressive advances, the testing revealed persistent issues across all platforms. These include physics inconsistencies (objects appearing and disappearing), aspect ratio limitations, facial changes in video editing, and unreliable audio generation. (37:33) Matt noted frustration that while image models can handle aspect ratio changes through outpainting, video models cannot yet perform similar transformations. Understanding these limitations is crucial for setting realistic expectations and planning workflows that work around current technical constraints.
Statistics & Facts
- Kling AI offers 33 video generations per month on their basic paid plan at under $9 monthly, making it significantly more economical than Google's Veo which costs $250 per month for unlimited access. (41:27)
- Google's Veo standard plan allows only approximately 10 video generations per month on their $20 monthly subscription, highlighting the cost barriers for regular usage of premium AI video tools.
- Matt has been using "monkey on roller skates" as his benchmark test prompt for over 2.5 years, dating back to the ModelScope release in 2022-2023, demonstrating the rapid evolution in AI video quality over this relatively short timeframe. (04:32)
