FastH3-Omni-NVFP4-QAD-Comfy: 16GB+ VRAM, File List

FastH3-Omni-NVFP4-QAD-Comfy: 16GB+ VRAM, File List

At a Glance

Item Value
Repository FastVideo/FastH3-Omni-NVFP4-QAD-Comfy
Family guide MiniMax-H3 guide (14 articles)
Publisher guide MiniMax: models and licenses
Published 2026-10-21
License minimax-h3-community-license-agreement
Formats safetensors
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

FastVideo has released “FastH3-Omni-NVFP4-QAD-Comfy", an audio-enabled video generation model that can be used directly in ComfyUI. This model accepts multiple references such as images, videos, and audio in addition to text prompts, and generates videos with synchronized 32kHz stereo audio based on them.

Based on the original model “FastVideo-FastH3-Omni-8-Step-V1", it is provided as a single safetensors file adapted for the ComfyUI environment. The quantization adopts the NVFP4 format using Quantization-Aware Distillation (QAD), supporting fast inference in just 8 sampling steps.

Specifications

The main specifications listed in the model card and the original model materials are as follows:

  • Input Specifications:
  • In addition to text prompts, accepts ordered reference inputs. – The upper limit for inputs as references is up to 9 images, up to 3 videos, and up to 3 audio clips, for a total of up to 12 items. – References must include at least one image or video. – Reference tags are used within the prompt to specify the input order.
  • Output Specifications:
  • Video framerate: 24 FPS – Audio: 32-kHz stereo audio synchronized with the video – Resoltuions and durations that can be generated: Depend on the runtime and available memory used (480p is standard in the local profile of the original model).
  • Recommended Inference Settings:
  • Number of sampling steps: 8 steps (8 Transformer evaluations) – Guidance scale (Guidance): 1.0 – Video shift: 12.0 – Audio shift: 3.0
  • Quantization Architecture & Hardware Requirements:
  • Adopts the NVFP4 QAD (Quantization-Aware Distillation) format. – NVIDIA Blackwell generation hardware is required for hardware acceleration of NVFP4 inference. For other GPU families, using BF16 or FP8 variants is recommended.
  • Components Required for ComfyUI Execution:
  • In addition to the model itself (for diffusion_models), compatible text encoders, video VAEs, and audio VAEs must be placed. – An audio VAE is required to output synchronized audio. Also, the text encoder side is required to support the reference vision function.
  • License Conditions:
  • According to the repository metadata, the “minimax-h3-community-license-agreement" applies.

Performance and Quality

Tables summarizing benchmark scores or quantitative evaluations are not published in this model card or the base model materials. However, the publishers have specified the following points regarding generation quality and behavior.

First, as a performance characteristic of the original model, despite an extremely small number of inference steps—8 Transformer evaluations (8 steps)—it can incorporate multiple multimedia references and generate 24 FPS video without breakdown, synchronized with 32kHz stereo audio.

On the other hand, as a note regarding quality, it is reported that the quantized output by NVFP4 QAD may have differences compared to the output results of the BF16 version of the original model. In addition, since the ComfyUI version is executed using its own model layout, sampler, and runtime kernel, the publishers note that the output quality and behavior may also differ from when executed on other runtimes such as FastVideo’s local engine or UniServe.

Strengths and Use Cases

FastH3-Omni-NVFP4-QAD-Comfy excels at integrating multiple visual and acoustic references in addition to text instructions to generate high-quality video and audio together along the references.

The expected use cases and strengths of this model are as follows:

  • Reference-to-video based on multimedia references:
    Beyond generation solely from text prompts, diverse materials such as images, videos, and audio can be referenced as input references. It is possible to combine up to 9 images, up to 3 videos, and up to 3 audio clips (up to 12 items in total). It is suited for advanced direction, such as specifying character appearances with images, instructing dynamic scene developments with videos, or referencing acoustic materials like background music and sound effects.
  • Simultaneous generation of stereo audio synchronized with video:
    While utilizing soundtrack information included in video references, it can simultaneously output 32-kHz stereo audio matching the movement of the video. It is suitable for use cases in video production where you want to quickly produce cohesive content, eliminating the hassle of adding audio separately after video generation.
  • Fast experimentation in visual workflows with ComfyUI:
    This model can be directly incorporated into node-based workflows using ComfyUI’s “MiniMax H3 Reference to Video" node. Because inference can be completed in a short process of 8 sampling steps, it is suited for tasks of repeatedly conducting generation experiments while changing combinations of parameters, prompts, and input references.

How It Differs from Similar Models

Compared to MiniMax-H3 based models introduced on this site in the past, there are clear differences in the flexibility of input references and the targeted execution environment.

In 音声付き動画生成モデル「FastH3 V2」のFP8およびNVFP4版公開, which introduced a model that directly generates audio-enabled video from text, video generation starting from text prompts was central. In contrast, the current FastH3 Omni differs significantly in supporting multimodal conditioning where up to 12 reference materials such as images, videos, and audio can be input in an ordered manner.

Also, in 音声付き動画生成モデル「FastH3 Trim」8ステップ版が公開, which aimed for structural weight reduction by removing transformer blocks, the trade-off between size reduction and quality through block removal was verified. In contrast, FastH3-Omni-NVFP4-QAD-Comfy adopts the NVFP4 format using Quantization-Aware Distillation (QAD) instead of block reduction, and is provided as a file format dedicated to ComfyUI, thereby balancing usability and reference control capability.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch)

Your VRAM Quantization File size Est. memory needed
16GB (RTX 5060 Ti 16GB / 4060 Ti 16GB, etc.) NVFP4 12.8GB 15.4GB

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

The publisher distributes this model as safetensors.

License — minimax-h3-community-license-agreement (Commercial use prohibited): MiniMax’s MiniMax H3 Community License. Use of the weights is limited to the EU, the UK, South Korea and the US (Sections II and V.4); elsewhere, including Japan, a separate license is required.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Our Own Measurements

Measurements We Did Not Take

This model’s license (minimax-h3-community-license-agreement) does not permit commercial use. Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).

Distributed Files

Weight files published in FastVideo/FastH3-Omni-NVFP4-QAD-Comfy, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
fasth3_omniref_qad_nvfp4_comfy.safetensors 13.77GB

How to Get It

This model is distributed as a single safetensors file on Hugging Face, and since it is not a gated model, it can be downloaded without prior license agreement procedures.

The file placement destinations and execution environment preparation procedures are as follows:

  • Placement of the Model Itself:
    Obtain the distributed fasth3_omniref_qad_nvfp4_comfy.safetensors and place it in the ComfyUI models directory.
ComfyUI/models/diffusion_models/fasth3_omniref_qad_nvfp4_comfy.safetensors
  • Preparation of Common Components:
    In addition to the model itself, you need to prepare a compatible text encoder, video VAE, and audio VAE. Obtain these from the FastH3-OmniRef-Comfy repository and place them in the following respective folders:
ComfyUI/models/text_encoders/
ComfyUI/models/vae/

An audio VAE is required to output audio synchronized with the video. Also, the text encoder needs to support the reference vision function, and it is guided to select the format according to the GPU you use.

  • Preparation of the ComfyUI Environment:
    Use the latest ComfyUI that natively supports MiniMax-H3, and use the “MiniMax H3 Reference to Video" node for conditioning. When referencing references within prompts, specify and write tags according to the order.

  • Notes on Runtime Format:
    This model is created in a layout and format dedicated to ComfyUI. When running it with FastVideo’s local inference engine, you must use FastVideo/FastH3-OmniRef-NVFP4-QAD, which is a model in the FastVideo format provided in a separate repository. Please select the checkpoint format that matches your execution environment.

Related Articles

What to Read Next

Sources