Train a MiniMax H3 Ref2VA LoRA with AI Toolkit — MiniMax H3 tutorial
Prepare a video-to-video dataset and run an auditable Ref2VA LoRA training experiment.
Audience: Developers with training experience, data, and a high-memory GPU.
Hardware: A high-memory NVIDIA GPU; use current AI Toolkit and H3 training docs for exact requirements.
Prerequisites
- Video data you are allowed to use
- A current AI Toolkit release
- The base-model license has been reviewed
Steps
- Define the target behavior as verifiable data pairs.
- Configure a Ref2VA LoRA from the tutorial instead of copying image-LoRA settings.
- Run a short training and fixed validation set before scaling up.
Caveats
- Training is compute-intensive; confirm dataset rights and model licensing first.
View original source · Ostris · 2026-08-23