FastH3-Omni-FP8-Comfy Audio-Synced Video Generation Model: 32GB+ VRAM

At a Glance
| Item | Value |
|---|---|
| Repository | FastVideo/FastH3-Omni-FP8-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-FP8-Comfy", an FP8 quantized video and audio generation model that can be executed directly in ComfyUI.
This model is capable of generating videos containing synchronized 32kHz stereo audio in just 8 sampling steps, by taking text prompts along with ordered image, video, and audio references. It is a conversion of the base model “FastVideo-FastH3-Omni-8-Step-V1" into ComfyUI’s native FP8 format, supporting lightweight local inference in the node-based ComfyUI environment.
Specifications
Specifications confirmed from the model card and base model documentation are as follows:
- Input Format: Supports text prompts as well as ordered reference inputs of images, videos, and audio. As a feature of the original model, 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). At least one image or video input is required for generation. When a video reference is input, its soundtrack information is also inherited.
- Output Specifications: 24fps video is output along with 32kHz stereo audio synchronized to the video. Supported resolutions and generated lengths depend on the execution environment and available memory.
- Sampling Settings: The recommended number of inference steps is 8 steps (8 Transformer evaluations). A guidance scale of 1.0, Video shift of 12.0, and Audio shift of 3.0 are specified.
- Prompt Syntax: Dedicated reference tags are used within the prompt to specify the reference order of source materials.
- Required Components: Operates by combining not only standalone diffusion model weights, but also a ComfyUI text encoder, video VAE, and an audio VAE which is essential for synchronized audio output. The text encoder requires support for reference-vision.
- Quantization Format: Distributed in FP8 format (single safetensors file format) using ComfyUI’s native quantization layout and kernels.
- License Conditions: The “minimax-h3-community-license-agreement" applies to this model and the base model.
Performance and Quality
In the model card, objective numerical scores from external benchmarks and comparison tables with existing models are not published.
However, regarding generation quality and the effects of quantization, the publishers provide the following notes on behavior:
- Differences Due to FP8 Quantization: Because this model utilizes ComfyUI’s native quantization layout and kernels, it is explained that differences may occur in the output results compared to the pre-quantization BF16 version (original model).
- Output Variations Across Runtimes: Since ComfyUI uses its own model layout, samplers, and runtime kernels, it is stated that output results may also differ from FastVideo’s native inference environment or execution results using UniServe.
- Quality Design During Fast Inference: Similar to the base model FastVideo-FastH3-Omni-8-Step-V1, the design of outputting video with synchronized audio in an extremely small number of steps—8 Transformer evaluations—is maintained. High-speed generation in just 8 steps is achieved through video and audio schedule settings and reference conditioning attention.
Strengths and Use Cases
FastH3-Omni-FP8-Comfy is suited for workflows that go beyond a single text instruction, providing multiple images, videos, and audio as complex references to quickly generate videos with a consistent worldview and soundscape in short steps.
Based on the model card descriptions and the design of the base model, the main strengths and intended use cases are as follows:
- Multimodal Reference Input: Allows the model to directly reference not only images, but also past video clips and audio materials. It can handle up to 9 images, 3 videos, and 3 audio items (up to 12 items in total) simultaneously, enabling conditioning that includes soundtrack information contained in video references.
- Video Generation with Synchronized Sound: Unlike methods where only video is generated and sound effects or BGM are added later, it simultaneously generates 24fps video and synchronized 32kHz stereo audio. It has strengths in scenarios where users want to obtain soundtracks tailored to movements and scenes within the video all at once.
- Visual and Interactive Prototyping via ComfyUI: Provided as a ComfyUI-native FP8 format, making it ideal for repeatedly testing material substitutions and prompt adjustments while combining with other nodes on a node-based GUI.
- Fast Generation Cycles: Because generation is completed in an extremely short sampling of 8 steps, it is suited for creators and engineers who want to quickly validate ideas on a local workstation without spending long hours of computing resources.
How It Differs from Similar Models
Specific differences from FastH3 derivative models introduced on our site in the past are as follows:
The article 音声付き動画生成モデル「FastH3 V2」のFP8およびNVFP4版公開 focused on text-to-video generation and performed NVFP4 or FP8 quantization for single-GPU environments based on the FastH3 V2 transformer itself, whereas this model, “FastH3-Omni-FP8-Comfy", differs significantly in supporting multimodal conditioning that can accept complex inputs of up to 12 reference materials such as images, videos, and audio. Additionally, unlike 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 lightweight operation through block reduction, this model focuses on adopting an FP8 layout optimized for ComfyUI’s native execution environment and realizing multi-reference capabilities rather than pruning transformer blocks.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch)
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 32GB (RTX 5090, etc.) | FP8 | 21.5GB | 25.8GB |
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-FP8-Comfy, listed by this site from the Hugging Face API. Sizes are the actual file sizes.
| File | Size |
|---|---|
fasth3_omniref_fp8_comfy.safetensors |
23.05GB |
How to Get It
FastH3-Omni-FP8-Comfy is available on Hugging Face in a single safetensors file format. The repository is accessible to everyone, and prior approval applications or Gated access procedures are not required for use.
When using the model, follow the steps below to place the files in the designated ComfyUI directory:
-
Place the Main Model:
Downloadfasth3_omniref_fp8_comfy.safetensorsand place it in ComfyUI’s diffusion models directory.text ComfyUI/models/diffusion_models/fasth3_omniref_fp8_comfy.safetensors -
Acquire and Place Common Components:
To run this model, not only the standalone diffusion model is needed, but compatible text encoders, video VAEs, and audio VAEs are also required separately. These can be obtained from FastH3 Omni Comfy and stored in their respective directories: – Text encoder:ComfyUI/models/text_encoders/(Select a format supported by your GPU and use one that supports the reference-vision feature) – VAE:ComfyUI/models/vae/(Introducing an audio VAE is mandatory to output synchronized audio) -
Prepare the ComfyUI Environment:
Use the latest ComfyUI build environment with native support for MiniMax-H3, and build the pipeline using the MiniMax H3 Reference to Video conditioning node.
Note that this checkpoint is in a format conforming to ComfyUI’s proprietary model layout and kernels. If you use FastVideo’s native inference runtime, you must use checkpoints provided separately in the FastVideo format rather than this model.
Related Articles
- FastH3-Omni-Comfy: 80GB+ VRAM, File List
- FastH3-Omni-NVFP4-QAD-Comfy: 16GB+ VRAM, File List
- FastH3-OmniRef-Comfy: File List
- FastVideo-FastH3-Trim-Comfy Text-to-Video and Audio Model: File List
What to Read Next
- Find models by VRAM (This model needs at least 32GB) → VRAM quick reference
- Explore the same model family → MiniMax-H3 family overview (14 articles, 4 converted builds)
- What FP8 mean and where to get this model → FP8 format guide and models / Quantization and model-format glossary
- Learn about the publisher → MiniMax: models, licenses and articles

