ComfyUI model file · Diffusion model (video + audio)

minimax_h3_ref2va_pruned_fp8_scaled.safetensors

MiniMax H3 Ref2VA pruned (fp8 scaled) · Reference-to-video diffusion model for MiniMax H3

This is MiniMax H3’s Ref2VA model, pruned and stored in fp8 scaled. It makes video with sound from your prompt plus reference images of characters, objects or places. Put it in ComfyUI/models/diffusion_models and load it with Load Diffusion Model.

  • MiniMax H3
  • Reference to video
  • 20.96 GB
  • fp8 scaled

Civitai names this download minimaxH3PrunedFp8_ref2vaPrunedFp8.safetensors. Same bytes. Comfy-Org suggests int8_convrot instead when your PyTorch supports it.

minimax_h3_ref2va_pruned_fp8_scaled.safetensors

Type
Diffusion model (video + audio)
Format
Safetensors · fp8 scaled
Size
20.96 GB (20,958,205,608 bytes)
SHA256
f86f2f79…4b392b8e9
ComfyUI folder
models/diffusion_models/
SHA256 and size from Hugging Face file metadata (Comfy-Org/MiniMax-H3), matching the Civitai API hash, checked by RunComfy on .

The short version

What is minimax_h3_ref2va_pruned_fp8_scaled.safetensors?

MiniMax H3 ships as two diffusion models: FL2VA for text, image and first-last-frame to video, and Ref2VA for reference-guided video. This file is the Ref2VA one. You give it up to several reference images and a prompt, and it keeps those characters or objects consistent through the clip while generating matching audio.

“Pruned” means Comfy-Org removed weights the model does not need at inference, which drops the full model from 66 GB to 40 GB in bf16. “fp8 scaled” halves it again to 20.96 GB. Comfy-Org recommends this fp8 file only if you cannot use the int8_convrot version.

Civitai’s listing of the same file (model 2832970) renames it minimaxH3PrunedFp8_ref2vaPrunedFp8.safetensors on download. The Civitai API hash matches the Hugging Face file.

Install

Download and install minimax_h3_ref2va_pruned_fp8_scaled.safetensors

Where it goes
ComfyUI/
└─ models/
   ├─ …
   └─ diffusion_models/
      └─ minimax_h3_ref2va_pruned_fp8_scaled.safetensors
terminal · Linux / macOS
$ cd ComfyUI/models/diffusion_models$ wget -O minimax_h3_ref2va_pruned_fp8_scaled.safetensors \$   https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_pruned_fp8_scaled.safetensors$ sha256sum minimax_h3_ref2va_pruned_fp8_scaled.safetensors# f86f2f79ebd2d76eb8eeb46091e83982e6ff51d255747e7b16e92834b392b8e9

Safetensors files hold tensors only and cannot run code. Still compare the SHA256: a failed download often leaves an HTML error page saved under the model’s filename.

Use

How to load minimax_h3_ref2va_pruned_fp8_scaled.safetensors in ComfyUI

Load Diffusion Model reads the diffusion model and outputs a MODEL for the sampler. The text encoder and VAE load in their own nodes.

Minimal graph. Click a node to open its reference page.

Works with

Files a MiniMax H3 video workflow needs

minimax_h3_ref2va_pruned_fp8_scaled.safetensors is one part of the stack. A workflow that reports a missing model usually needs these too. Use the Ref2VA diffusion model instead of FL2VA for reference-to-video workflows.

FileRoleFolderSizeDownload
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensorsText encoder (Qwen3-VL 32B)models/text_encoders/15.69 GB Get
minimax_h3_video_vae_fp16.safetensorsVideo VAEmodels/vae/5.21 GB Get
minimax_h3_audio_vae_fp32.safetensorsAudio VAEmodels/vae/605 MB Get
minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensorsTurbo LoRA, 4 steps(optional)models/loras/1.96 GB Get

Run it

ComfyUI workflows that can use minimax_h3_ref2va_pruned_fp8_scaled.safetensors

We match workflows by the model files they load: this exact file first, then other versions of it, then workflows built for the same use, then other models from the same family.

  1. Uses this exact file1
  2. Uses another version of this file19
  3. Built for reference-to-video6

Uses this exact file

1 workflow

These workflows already load minimax_h3_ref2va_pruned_fp8_scaled.safetensors. Open one and run it as is.

Uses another version of this file

19 workflows

These workflows load a different precision or format of the same model. Select minimax_h3_ref2va_pruned_fp8_scaled.safetensors in their Load Diffusion Model node to use this version.

Built for reference-to-video

6 workflows

These workflows load another reference-to-video model in the same folder, or they are reference-to-video workflows that already have a Load Diffusion Model node. Select minimax_h3_ref2va_pruned_fp8_scaled.safetensors there.

Compare

Same file, other names, and alternatives

FileRelationSizeBest for
minimax_h3_ref2va_pruned_fp8_scaled.safetensorsThis page · pruned fp8 scaled20.96 GBGPUs without int8_convrot support
minimax_h3_ref2va_pruned_int8_convrot.safetensorsSame model · int8 convrot20.97 GBPreferred by Comfy-Org on PyTorch cu130
minimax_h3_ref2va_pruned_w6a8.safetensorsSame model · w6a815.98 GBLess VRAM
minimax_h3_ref2va_pruned_bf16.safetensorsSame model · bf1640.23 GBHighest fidelity, large GPUs
minimax_h3_fl2va_pruned_int8_convrot.safetensorsSibling model · FL2VA20.97 GBText, image and first-last-frame to video

Troubleshoot

Common errors and fixes

Value not in list: unet_name: 'minimax_h3_ref2va_pruned_fp8_scaled.safetensors' not in [...]

Why: The workflow asks for this exact filename, but ComfyUI cannot find it in models/diffusion_models. Usually the file is missing, sits in a different folder, or was saved under another name such as minimaxH3PrunedFp8_ref2vaPrunedFp8.safetensors.

Fix: Move the file into ComfyUI/models/diffusion_models/, rename it to match the workflow (or pick it again in the dropdown), then refresh the node list or restart ComfyUI.

torch.OutOfMemoryError: CUDA out of memory

Why: The model alone is 21 GB, before the text encoder, VAEs and activations for long clips.

Fix: Lower the resolution or clip length, enable model offloading, or switch to the w6a8 file. A 48 GB or larger cloud GPU runs it at full settings.

SafetensorError: Error while deserializing header

Why: The download is incomplete, or the browser saved an HTML sign-in or error page under the model’s filename.

Fix: Check that the file is exactly 20,958,205,608 bytes and that its SHA256 matches the value above. Download it again from Hugging Face · Comfy-Org/MiniMax-H3 if not.

Rather skip the setup?

Run MiniMax H3 Taomate 3-Step Acceleration in your browser on a cloud GPU. Nothing to download.

Generate video with MiniMax H3 online

FAQ

minimax_h3_ref2va_pruned_fp8_scaled.safetensors FAQ

Ref2VA or FL2VA, which one do I need?

Ref2VA when you want specific characters or objects from reference images to appear in the video. FL2VA for text to video, image to video, and first-last-frame to video. Many setups keep both.

Is minimaxH3PrunedFp8_ref2vaPrunedFp8.safetensors a different model?

No. It is Civitai’s download name for this file. The SHA256 f86f2f79ebd2d76eb8eeb46091e83982e6ff51d255747e7b16e92834b392b8e9 is the same on both sites.

Where do I put minimax_h3_ref2va_pruned_fp8_scaled.safetensors in ComfyUI?

In ComfyUI/models/diffusion_models/. Then refresh the node list or restart ComfyUI so Load Diffusion Model lists the file.

Which ComfyUI node loads minimax_h3_ref2va_pruned_fp8_scaled.safetensors?

Load Diffusion Model. It is built into ComfyUI, so no custom nodes are needed.

Can I use minimax_h3_ref2va_pruned_fp8_scaled.safetensors commercially?

Yes, within the MiniMax H3 Community License. Commercial products must show “MiniMax H3” in their interface, and companies with more than US$20 million in yearly revenue need written permission from MiniMax first. The license also limits use to its listed territories, so read it before shipping.

Sources: Comfy-Org/MiniMax-H3 on Hugging Face · MiniMaxAI/MiniMax-H3 (original weights) · ComfyUI MiniMax H3 tutorial · Civitai model 2832970

RunComfy
Copyright 2026 RunComfy. All Rights Reserved.

RunComfy is the premier ComfyUI platform, offering ComfyUI online environment and services, along with ComfyUI workflows featuring stunning visuals. RunComfy also provides AI Models, enabling artists to harness the latest AI tools to create incredible art.