MiniMax-H3-TAE

MiniMax-H3-TAE

A small, quickly trained decoder that gives live image previews while MiniMax-H3 video generates in ComfyUI. It gives a cleaner preview than the plain latent2rgb option.

Made by Kijai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: MiniMax-H3-TAE is a small 2D VAE that Kijai trained in a hurry for the MiniMax-H3 video model. It turns the model's latent frames into rough pictures, so you can watch a generation as it runs. The author calls the result modest, yet says it looks better than the default latent2rgb preview.

Who it suits: It suits people who run MiniMax-H3 in ComfyUI and want a clearer preview while a video renders. It is a helper for one narrow job, not a model that makes video by itself.

Title: Who might use it
Industry | Who uses it | How they use it | Example
AI video hobby | ComfyUI users | Swap in the decoder for the preview | Watch a MiniMax-H3 run and see whether it drifts
Workflow building | Node and workflow authors | Compare preview decoders in one graph | Put it next to latent2rgb and judge the difference
Creative production | Video artists | Check sampling progress by eye | Glance at the preview before waiting for the final clip

If you never run MiniMax-H3, or you only care about the finished video, you can skip it.

Where to use it: The author says it works only with one node for now, so ComfyUI is the place.

• ComfyUI, through the ModelPreviewOverride node in the ComfyUI-KJNodes pack
• Windows, macOS or Linux, with Python, to download the files

Title: What it does in practice
Use case | Who | What they get back
Live preview of a MiniMax-H3 run | ComfyUI users | A rough picture of each step while sampling
Preview comparison | Workflow tinkerers | A side by side look against latent2rgb

How to set it up: You need a few things in place before the files are any use.

• Python with pip, to fetch the files from Hugging Face
• A ComfyUI install that has the ComfyUI-KJNodes pack
• The MiniMax-H3 model you want to preview

1. Install the Hugging Face download tool.
2. Download this repository to your computer.
3. Add the ModelPreviewOverride node to your workflow, since the card says the decoder works only there.

pip install -U huggingface_hub
hf download Kijai/MiniMax-H3-TAE

How to use it: Run your MiniMax-H3 workflow with the ModelPreviewOverride node in place. The preview you see while sampling then comes from this decoder, not from latent2rgb. The model card has no further settings, so the node itself is the only control to learn.

hf download Kijai/MiniMax-H3-TAE

Title: The pieces
Tool | What it does
ModelPreviewOverride | The ComfyUI-KJNodes node that swaps in this decoder for previews
hf download | Fetches the repository files to your computer

Example: You start a MiniMax-H3 video in ComfyUI with the preview node switched on. Instead of the flat colour smear that latent2rgb gives, you see a rough but readable picture of the frames forming. You can tell early whether the run is going the way you wanted.

For newcomers and experts: If you are new, think of it as a cheap viewfinder. Install the files, switch on the preview node in a MiniMax-H3 workflow, and just watch what changes. Nothing else in your workflow needs to change.

If you know the field, this is a tiny autoencoder trained to decode latents quickly, in the spirit of the TAEHV family. It trades fidelity for speed, so it is a preview aid and not a final decoder. The author also points to a properly trained version from madebyollin that works with the same node.

MiniMax-H3 latents arrive while sampling
The preview node hands them to the tiny decoder
The decoder returns a rough picture
You watch it in ComfyUI

Good to know: Quality is deliberately modest, and the author says a better-trained decoder now exists. It works only with the ModelPreviewOverride node in ComfyUI-KJNodes at the moment.

• License: apache-2.0
• Cost to run: none named, the files download free from Hugging Face
• Data it reaches: the files you download, and the frames inside your own ComfyUI run
• Not for: final video output, or any setup without the ModelPreviewOverride node

On Hugging Face, 13 November 2026

• Downloads in the last 30 days: 0
• Likes: 195
• License: apache-2.0

Links

• Model: https://huggingface.co/Kijai/MiniMax-H3-TAE
• Files: https://huggingface.co/Kijai/MiniMax-H3-TAE/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Kijai/MiniMax-H3-TAE

Then follow the usage in its model card. Nothing in this kit runs by itself.