MiniMax's Hailuo 02 Outperforms Google Veo 3 in User Benchmarks at Lower Video Costs

MiniMax has unveiled the second generation of its video AI model, Hailuo 02, with significant improvements in both performance and cost efficiency. The new model leverages an innovative architecture called Noise-aware Compute Redistribution (NCR), which MiniMax claims enhances training and inference efficiency by a factor of 2.5.

Key Technical Upgrades

• Noise-aware Compute Redistribution (NCR) Architecture:

  • Early training stages: Introduces heavy artificial noise, compressing videos to reduce computational load.
  • Later training stages: Processes clearer videos at full resolution for fine-tuning and accuracy.

• Parameter Count and Training Data:

  • Hailuo 02 has three times more parameters and four times more training data compared to its predecessor.
  • The company emphasizes improvements in data quality and diversity but hasn't disclosed exact parameter counts or dataset sizes.

Performance Gains

Hailuo 02 demonstrates notable advancements in handling complex prompts and simulating physical processes. MiniMax asserts that it is currently the only model capable of accurately generating intricate scenes such as gymnastics routines.

Video Quality and Variants

The new model offers three variants:

• 768p for six seconds
• 768p for ten seconds
• 1080p for six seconds

These options represent a significant upgrade from the previous model, which was limited to 720p, six-second videos at 25 fps.

Benchmark Performance

In the Artificial Analysis Video Arena benchmark, where users rate videos generated by competing AI models:

• Hailuo 02 secured second place in the image-to-video category.
• It trailed slightly behind Bytedance's Seedance but significantly outperformed Google's highly anticipated Veo 3.
• Notably, this version of Veo 3 does not support audio, which is a crucial feature for many applications.

In user benchmarks, Hailuo 02 continues to outperform Google Veo 3, even though Veo supports native audio generation. This suggests that the quality and efficiency gains from MiniMax's NCR architecture are substantial enough to offset the lack of audio in these tests.

User Adoption

Since its demo launch in August last year, over 3.7 billion videos have been created using the Hailuo platform, according to MiniMax. The company attributes this rapid adoption to the model's enhanced capabilities and cost-effectiveness.

Conclusion

MiniMax's Hailuo 02 represents a significant leap forward in video AI technology, combining advanced architectural innovations with practical performance gains. Its ability to outperform more established models like Google Veo 3, while maintaining lower costs, makes it an attractive option for developers and content creators alike.