Official

minimax / hailuo-02

Hailuo 2 is a text-to-video and image-to-video model that can make 6s or 10s videos at 720p (standard) or 1080p (pro). It excels at real world physics.

  • Public
  • 15K runs
  • Commercial use
Iterate in playground
Run with an API
Playground API Examples README

MiniMax Hailuo 02

MiniMax Hailuo 02 is a video generation model designed for high-quality video synthesis from text and image inputs. The model represents a significant architectural advancement over its predecessor, featuring improved efficiency and enhanced capabilities for complex video generation tasks.

Standard and Pro

There are two models, standard and pro:

  • 720p (standard)
  • 1080p (pro)

The pro model produces higher quality videos with better physics and coherency.

Architecture

The model is built on the Noise-aware Compute Redistribution (NCR) framework, a novel architecture that optimizes computational efficiency during both training and inference. This approach redistributes compute resources based on noise levels in the generation process, achieving 2.5x improvement in training and inference efficiency compared to conventional architectures at similar parameter scales.

Model Specifications

  • Parameters: 3x larger than predecessor model
  • Training Data: 4x larger dataset compared to previous version
  • Native Resolution: 1080p output capability
  • Output Formats:
  • 768p at 6 seconds
  • 768p at 10 seconds
  • 1080p at 6 seconds

Key Capabilities

Instruction Following

The model demonstrates state-of-the-art performance in interpreting and executing complex text prompts, accurately translating detailed instructions into corresponding video content.

Physics Simulation

Advanced physics modelling capabilities enable realistic representation of complex physical interactions and movements, including challenging scenarios like gymnastics and intricate motion sequences.

Multi-modal Input

Supports both text-to-video (T2V) and image-to-video (I2V) generation workflows, allowing for flexible content creation approaches.

Technical Performance

  • Efficiency: 2.5x improvement in computational efficiency over comparable models
  • Quality: Native 1080p generation capability
  • Stability: Enhanced alignment and generation success rates
  • Versatility: Handles complex scenarios requiring precise physics simulation

Training Details

The model was trained on a significantly expanded dataset featuring improved data quality and diversity. Training incorporated user feedback and usage patterns from the predecessor model to optimize performance on real-world video generation tasks.

Limitations

  • Generation speed continues to be an area for improvement
  • Model alignment and stability, while enhanced, remain ongoing development focuses
  • Current capabilities are primarily focused on T2V and I2V tasks

Model Card

Model Type: Video Generation
Architecture: Noise-aware Compute Redistribution (NCR)
Input Modalities: Text, Image
Output Modality: Video
Maximum Resolution: 1080p
Maximum Duration: 10 seconds

Development Status

This model represents an active development effort with ongoing improvements planned for:

  • Generation speed optimization
  • Enhanced model alignment and stability
  • Expansion beyond current T2V and I2V capabilities

Privacy policy

Data from this model is sent from Replicate to MiniMax.

Check their Privacy Policy for details:

https://intl.minimaxi.com/protocol/privacy-policy

Terms of Service

https://intl.minimaxi.com/protocol/terms-of-service