H3-World: Interactive Image-to-Video Model Based on MiniMax-H3

A model called DANNY621/H3-World has been published on HuggingFace with an image-to-video task. The author describes it as the first interactive world model based on MiniMax-H3: given an initial frame and keyboard commands, it generates video with consistent character and camera motion. The model is released under the apache-2.0 license, the largest weights file is step-10000.safetensors at 0.1 GB, and the weights format is safetensors. The checkpoint is a LoRA adapter with 65.6M parameters on top of the base MiniMax-H3 model and requires its weights and a directional attention patch to work.

What it means

Our reference on video generators includes several open image-to-video and text-to-video models that can be compared with this new release. LTX-Video from Lightricks is a video generator with an image-to-video task, open source under Apache-2.0, and the largest weights file at 26.6 GB. The authors claim a need for only 1 GB of VRAM and support for MPS on macOS. Against this backdrop, H3-World looks noticeably lighter in terms of weights: its LoRA file takes up 0.1 GB, but running it still requires the base MiniMax-H3 model, whose weight is not listed in the H3-World fact sheet. CogVideoX from Zhipu AI handles a text-to-video task, has open source under Apache-2.0, and the largest weights file at 9.2 GB. The platform as described by the author is CUDA. H3-World differs from it both in task (image-to-video with control) and approach: instead of a full model, it is a LoRA adapter on top of a third-party base. Pyramid Flow is a text-to-video generator with open source under MIT, the largest weights file at 7.8 GB, and claimed operation on less than 8 GB of VRAM. Like CogVideoX, it is a full model, whereas H3-World is a parameter-efficient adapter. H3-World stands out not for its size but for its approach: interactive control of generation via keyboard commands and binding instructions to latent intervals through directional attention. For practical use, the base MiniMax-H3 model and the official repository with the patch are required — without them, the LoRA checkpoint will not reproduce the claimed behavior.