Introduction

The one we've been waiting on is finally out. FastH3 Omni, Hao AI Lab's distilled version of MiniMax H3 for reference-to-video, is now open weights. FastVideo announced it on October 20 (19:22 UTC, 03:22 MYT on October 21) with the training code included, and ComfyUI confirmed native support about an hour later.

It's a video + audio model: you give it a prompt plus ordered image, video and audio references, and it returns one clip with synchronized stereo sound in eight sampling steps. You can run it today in ComfyUI, in Hao AI Lab's FastVideo runtime, or straight in the browser on a free Hugging Face Space.

What shipped

  • Model: FastH3 Omni (also called FastH3 OmniRef), an 8-step distilled student of MiniMax H3 built by Hao AI Lab's FastVideo team
  • Partners named in the launch: NVIDIA, Nuva Lab and Reactor
  • Mode: reference-to-video with audio (Hao AI Lab calls it Ref2VA)
  • Sampling: 8 transformer passes at guidance 1.0
  • Output: video with synchronized 32 kHz stereo audio at 24 fps
  • Clip length in FastVideo: 5, 10 or 15 seconds (124, 243 or 362 frames)
  • License: MiniMax-H3 Community License Agreement, same as the base model
  • Training code: included, per the FastVideo announcement

Hao AI Lab's headline claim is real-time generation 12x faster on NVIDIA Blackwell, with NVFP4 and FP8 builds aimed at DGX Spark and RTX cards.

How references work

This is the fun part. You can feed it up to nine images, three videos and three audio clips, with no more than twelve references in total, and at least one image or video is required. Video references bring their soundtracks along.

Inside the prompt, you point at each input by its order with tags like <Picture 1>, <Video 1> and <Audio 1>. So a character sheet, a dance clip and a voice sample can all land in one shot. Here's a starter prompt to adapt:

<Picture 1> walks onto a rain-soaked neon rooftop at night and performs the dance moves from <Video 1>, the camera slowly circling her at waist height, puddles reflecting magenta and teal signs, while she sings along to <Audio 1> with matching lip movement, warm gold rim light, cinematic shallow depth of field

Running it in ComfyUI

Use a current ComfyUI build with native MiniMax H3 support and the MiniMax H3 Reference to Video conditioning node. The model is not in a tagged release note of its own yet (the latest tag is still v0.42.1), so update to the newest build first.

Pick one diffusion model:

  • BF16: fasth3_omniref_8step_pruned_bf16.safetensors, 41.2 GiB, from FastH3-Omni-Comfy
  • INT8 ConvRot: fasth3_omniref_8step_pruned_int8_convrot.safetensors, 20.7 GiB, same repo
  • FP8: fasth3_omniref_fp8_comfy.safetensors, 21.5 GiB, from FastH3-Omni-FP8-Comfy
  • NVFP4 QAD: fasth3_omniref_qad_nvfp4_comfy.safetensors, 12.8 GiB, from FastH3-Omni-NVFP4-QAD-Comfy

Then add one Qwen3-VL 32B text encoder (BF16 48 GiB, INT8 25.3 GiB, or NVFP4 AWQ 14.6 GiB), one video VAE (FP16 4.85 GiB or INT8 2.62 GiB) and the audio VAE (0.56 GiB), which you need for synced sound. All of them live in the text_encoders/ and vae/ folders of the FastH3-Omni-Comfy repo. The NVFP4 encoder needs Blackwell, so pick BF16 or INT8 on older cards.

Recommended settings from the model card:

  • 8 steps, guidance 1.0
  • Video shift 12.0, audio shift 3.0

The card also warns that ComfyUI uses its own sampler and kernels, so results can differ a little from FastVideo.

Running it locally with FastVideo

For the fastest local path, FastVideo pairs the NVFP4 QAD checkpoint with device profiles:

  • DGX Spark (128 GB unified): 480p, everything stays resident
  • RTX PRO 6000 (96 GB): 768p, everything stays resident
  • RTX 5090 (32 GB): 480p, text encoder offloaded, needs a precomputed AdaLN table
  • RTX 4090 (24 GB): 480p on the FP8 weights, streamed layer by layer since there's no FP4 hardware

QAD means quantization-aware distillation, so these weights were trained to run in NVFP4 rather than just squashed afterward. The example from the FastVideo guide looks like this:

hf download FastVideo/FastH3-Omni-NVFP4-QAD --local-dir ./FastH3-Omni-NVFP4-QAD
python examples/inference/basic/basic_fasth3_omniref_pdd.py \
  --model-path ./FastH3-Omni-NVFP4-QAD \
  --image reference.png --audio voice.wav \
  --prompt "a person on a sofa, lit by a warm lamp, speaking to camera" \
  --profile rtx-5090

The local inference guide covers the AdaLN table, offload settings and the optional NVFP4 VAE decoders we covered yesterday. Data-center users on B200, GB200, RTX PRO 6000 or Hopper can serve the full BF16 checkpoint (about 65 GiB of transformer weights) through Hao AI Lab's UniServe.

Original Source

Hao AI Lab's launch post:

The FastH3 Omni collection holds every variant, the BF16 model card lists inputs, outputs and the license, and the FastH3-Omni Space lets you try it without a GPU.

Conclusion

Earlier this week FastH3 Omni was a private repo we kept checking. Now it's open, it reads pictures, clips and sound in one prompt, and it runs on a single consumer Blackwell card. Grab a character image and a voice line, try the Space first, then bring it home to ComfyUI 💜

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&amp;utm_campaign=gen&amp;utm_content=fasth3-omni-open-weights-reference-to-video-comfyui

—Aurelia ♡