Fast H3 Omni speeds up MiniMax video generation by up to 12x

Fast H3 Omni is an open-source acceleration framework built for MiniMax H3, aimed at turning sluggish video generation into a snappy, production-friendly workflow.

Instead of focusing purely on generic text-to-video, it is purpose-built for reference-to-video workflows. Feed it reference images of a character or a specific environment, and it keeps identity and scene details consistent across every generated frame without drifting off-model.

How it gets so fast

Co-developed by FastVideo and Nuva Lab alongside NVIDIA's FastGen team, the system cuts generation down from the base model's 50 sampling steps to just 8 steps.

That architectural trim delivers massive gains. In benchmark setups running across eight high-end GPUs, inference latency dropped by up to 12x compared to vanilla MiniMax H3. Even with heavier reference inputs, the speedup holds up: feeding three reference images yields about a 5.8x acceleration, while an input reference video runs roughly 6.98x faster.

It also supports synchronized sound. By wiring in a dedicated audio VAE, you can produce synced video and audio outputs simultaneously using simple reference tags like Picture, Video, and Audio directly inside your prompts.

Hardware and deployment

The full suite of weights is already live on Hugging Face under the MiniMax-H3 Community License Agreement.

FastVideo released several quantized variants suited for different setups, spanning BF16, INT8 ConvRot, FP8, and NVFP4 QAD, along with a build optimized specifically for Apple Silicon machines.

For most creators, the smoothest path is via ComfyUI. The FastH3-Omni-Comfy repository packs the required text encoders and VAEs, letting you run native reference-to-video generation with an 8-step sampler.

If you are running dedicated enterprise gear like GB200 or RTX PRO 6000 systems, you can tap the BF16 release through UniServe. For local power users on an RTX 5090 or DGX Spark, FastVideo provides recipe guides for NVFP4, while RTX 4090 owners can leverage FP8 weights with layerwise offloading.

Why it matters

If you build short-form narrative videos, game cinematics, or storyboard concepts, keeping faces consistent across cuts has always been the single biggest bottleneck. Condensing MiniMax H3 to an 8-step run makes high-fidelity reference pipelines practical outside massive compute clusters.

In our view, making this work inside ComfyUI with aggressive 4-bit and 8-bit quantization is the right playbook. It brings one of the strongest open video architectures straight to creator workstations, setting a high bar for open-source media generation tools.