MiniMax H3 TAE: Tiny 2D Autoencoder for Fast Previews in ComfyUI

Kijai released a quickly trained 2D tiny autoencoder (TAE) for MiniMax H3, giving ComfyUI users a fast way to preview H3 video latents during generation (repository). It is not a full-quality VAE, but it beats latent2rgb for preview purposes, which is useful when iterating on prompts and settings without waiting for the full decode.

This is a preview tool, not a production VAE: Kijai describes the result as "not the greatest outcome," but better than a plain latent-to-RGB preview. For a higher-quality option, madebyollin's properly trained taeh3.safetensors (taehv repo) also works with the same node.

A frame from Kijai's TAE preview demo

ComfyUI usage

The TAE currently works through the ModelPreviewOverride node in ComfyUI-KJNodes. Place the TAE weights where the node expects them, then point ModelPreviewOverride at the H3 video VAE slot to preview latents with the tiny decoder instead of the raw RGB conversion.

Kijai's TAE preview demo video

Availability

• TAE weights: Kijai/MiniMax-H3-TAE on Hugging Face (Apache-2.0)
• Higher-quality alternative: madebyollin/taehv — taeh3.safetensors
• Required node pack: ComfyUI-KJNodes