FastH3 Omni

FastH3 Omni is an open-weight, distilled version of MiniMax H3 for reference-guided video with sound, released by FastVideo (Hao AI Lab) on October 20, 2026. It turns a prompt plus reference images, video clips and voice into a clip with synchronized audio in 8 steps instead of 50, and FastVideo reports 5.6–11.9× lower latency than the base model. GoEnhance does not host FastH3 Omni; the console below runs the full MiniMax H3 reference-to-video model it was distilled from.

Reunion in the Lantern Alley

Corgi on the Cliff Path

Fox in the Snow

Vintage Camera Ad

MiniMax H3

Reference images

Add 1–4 images: a person, a pet, a product or a place. In your prompt, call them @[image1], @[image2]… in this order.

Prompt

Example prompts

By uploading a file, you agree to GoEnhance's Terms of Use and Privacy Policy

What FastH3 Omni Brings

8 Steps Instead of 50

FastH3 Omni is a few-step student of MiniMax H3: each clip takes eight transformer evaluations, down from 50 in the base model. Across 27 matched workloads on eight NVIDIA GB200 GPUs, FastVideo measured 5.6–11.9× lower end-to-end latency. The roughly 12× headline is the best case, a 15-second clip from one short reference video.

References for Faces, Scenes, Motion and Voice

Unlike the earlier FastH3, which handled text-to-video and image-to-video only, Omni keeps H3's reference-to-video mode. You can mix reference images, video clips and a voice reference in one prompt; FastVideo tested setups from a single image up to 9 images plus 3 clips.

Video and Sound in One Pass

Like MiniMax H3, FastH3 Omni generates picture and synchronized stereo audio together. FastVideo's examples run at 24 fps in 5, 10 and 15-second lengths.

Runs on Blackwell and RTX GPUs

Besides data-center GPUs, FastVideo ships NVFP4 weights trained for Blackwell, with recipes for DGX Spark and RTX 5090 that default to 480p, and an FP8 recipe for RTX 4090 with layer-by-layer offloading.

Open Weights, Training Code and ComfyUI

The weights are on Hugging Face in the FastVideo/fasth3-omni collection, including FP8 and NVFP4 packages for ComfyUI, and the training recipe is in the FastVideo repository. FastVideo says training took about 7,800 GPU hours on 64 NVIDIA B300s.

FastH3 Omni vs FastH3 vs MiniMax H3

Feature	FastH3 Omni	FastH3 (earlier release)	MiniMax H3 on GoEnhance
What it is	FastVideo's distillation of H3 for reference-to-video	FastVideo's distillation of H3 for text and image to video	The full model from MiniMax
Released	October 20, 2026	August 2026 (preview)	Available now
Sampling	8 steps	8 steps in ComfyUI (the first preview used 4)	Full schedule (50 steps in the open release)
Reference images, video and voice	Yes	No	Yes: up to 9 images, 1 video, 3 audio clips
Native audio	Yes	Yes	Yes
Where to run it	Your own GPUs, ComfyUI or FastVideo	Your own GPUs, ComfyUI or FastVideo	In the browser, no setup

How to Make Reference Videos Like FastH3 Omni on GoEnhance

Upload Your References

Add up to 9 images for characters, products and places. You can also add one video clip for motion and up to 3 audio clips for voice or sound, each 2–15 seconds.

Tag Them in the Prompt

Refer to each reference by its tag, such as @[image1], and describe the action, the camera and what we hear. Put spoken lines in quotes.

Choose Length, Shape and Quality

Pick 5 to 15 seconds, an aspect ratio from 16:9 to 9:16, and 720p or 1080p.

Generate and Refine

MiniMax H3 renders the clip with native sound. Adjust the prompt or swap a reference and run it again.

Try MiniMax H3 Reference to Video

You may want to know

What is FastH3 Omni?

FastH3 Omni is a distilled version of MiniMax H3 released by FastVideo, the video generation team at Hao AI Lab. It is trained to reproduce H3's reference-to-video mode, where you guide a clip with images, video and voice, in 8 steps instead of 50.

How much faster is FastH3 Omni than MiniMax H3?

FastVideo reports 5.6–11.9× lower end-to-end latency than the base model across 27 workloads on eight GB200 GPUs. For example, a 5-second clip from one image dropped from about 35 seconds to under 6, and a 15-second clip from 9 images and 3 videos dropped from about 7 minutes to about 1. Times on a single consumer GPU will be longer.

Is the quality the same as MiniMax H3?

Not necessarily. FastVideo shares side-by-side examples but says they are qualitative and not evidence of matching output. For the earlier FastH3, ComfyUI's guide notes that speed is traded for some motion and audio fidelity and recommends the base H3 when quality matters most.

How is FastH3 Omni different from FastH3?

The earlier FastH3 covered text-to-video, and in ComfyUI image-to-video, but its reference mode was not distilled. FastH3 Omni adds the reference-to-video path, so images, video clips and voice can steer the result.

Can I run FastH3 Omni on my own GPU?

Yes, with the right setup. FastVideo provides recipes for DGX Spark and RTX 5090 using NVFP4 weights, both defaulting to 480p, and an FP8 recipe for RTX 4090 with offloading. The bundled NVFP4 text encoder needs a Blackwell GPU.

Does FastH3 Omni work in ComfyUI?

Yes. FastVideo publishes ComfyUI packages of FastH3 Omni in FP8, NVFP4 and BF16/INT8 in its Hugging Face collection.

Is FastH3 Omni open source?

The weights are published on Hugging Face under FastVideo/fasth3-omni and the training code is in the FastVideo repository. Read the license on the model card before using it in a product.

Can I use FastH3 Omni on GoEnhance?

Not at the moment. The console on this page uses the full MiniMax H3 reference-to-video model, which takes the same kinds of references (up to 9 images, a video clip and up to 3 audio clips) and runs in your browser with no GPU. It costs 10 tokens per second at 720p and 14 at 1080p and is available to members.

Were the examples on this page made with FastH3 Omni?

No. All four examples were generated with MiniMax H3 reference-to-video on GoEnhance, the same model the console uses, at 480p and 5 seconds. The reference images are shown on each example.

Direct a Scene from Your References

Combine characters, places, a motion clip and a voice, and let MiniMax H3 turn them into a video with sound. No GPU or setup needed.