Imperial College releases LoRA adapter that makes MiniMax-H3 a real-time video generator

A research team at Imperial College London has published a LoRA adapter that turns the MiniMax-H3 video model into a causal streaming generator, meaning the model produces each new segment based only on previously generated content rather than sampling an entire bidirectional clip at once. The adapter, developed under the RAVEN project, is available as an early preview on Hugging Face, with the full code hosted in the project’s GitHub repository.

In standard video models, generation proceeds by denoising an entire frame sequence simultaneously, which requires holding the full clip in memory and limits generation length. RAVEN replaces this with causal extrapolation: each new segment is derived solely from its predecessors, analogous to token-by-token text generation. According to Yanzuo Lu, Ronglai Zuo and Jiankang Deng, the adapter does not increase computational requirements beyond the base model and may even reduce them, because each forward pass of the DiT processes only a single segment.

The authors stress that the release is a preview version: the adapter remains undertrained and texture details are still limited. Current work prioritises improving texture quality, while capabilities such as long-range generation via extensions like capped RoPE have not yet been explored, although the architecture theoretically supports them. The method is described in the preprint “RAVEN: Real-time Autoregressive Video Extrapolation with Consistency-model GRPO” on arXiv (2605.15190), which has not undergone peer review.

The adapter is released under the community license of MiniMax-H3, and users must comply with the base model’s licensing terms, the repository, and its dependencies. Running the system requires separately downloading the MiniMax-H3 components (tokenizer, DiT, text encoder, video VAE, audio VAE) and updating the paths in the supplied configuration files. The recommended execution environment consists of a single node with eight GPUs using 4-way unified parallelism and an FSDP shard size of 8.

Although limited, the current version validates the full training-to-inference pipeline of RAVEN on MiniMax-H3, from training through generation. The codebase includes training implementations in the projects/minimax_h3/meta_models/ directory, and community contributions are welcomed. The team’s next step is to accelerate inference to achieve genuine real-time generation while continuing to enhance texture quality.