Alibaba PAI released ControlNet-Union for the MiniMax-H3 model

The Alibaba PAI team introduced MiniMax-H3-Fun-Controlnet-Union — a specialized checkpoint for controlling video generation in the MiniMax-H3 model. The tool allows users to define the structure and composition of a video sequence through various control methods.

The solution supports six key types of input data: Canny, Depth, HED, MLSD, Pose, and inpainting. This enables more precise customization of generation based on contours, depth maps, skeletal animation, or selected areas for editing within a frame.

This release is classified as a legacy version. Developers officially recommend using the updated version 2.0 to achieve more stable and high-quality results when working with video streams. The tool is available for integration into existing video generation pipelines based on the MiniMax architecture.

Implementing such controllers allows the video creation process to be translated from random generation mode to a manageable creative process. Using ControlNet provides the ability to maintain consistency of movements and objects, which is critical for professional work with neural network video content.