LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler

Neural latent upscaler for Minimax H3 (24ch). Bypasses costly 5B-param VAE decode/encode. Upscale low-res latents directly, then refine. Accelerates high-res video gen, outperforms naive interp.

What it solves

This project provides a neural latent upscaler specifically designed for Minimax H3 video generation. It eliminates the need for the slow and computationally expensive process of decoding latents to pixels, upscaling them, and then re-encoding them back into latents. By upscaling directly in the latent space, it significantly accelerates high-resolution video generation while avoiding the ghosting and double-image artifacts typically caused by basic bilinear or bicubic interpolation.

How it works

The tool is implemented as a set of ComfyUI nodes that use a trained neural network to increase the spatial resolution of 24-channel Minimax H3 VAE latents. It offers two different architectural backbones:

  • 2D Variant: A lightweight and fast option using a 2D ResBlock backbone with Temporal 3D-Conv layers for temporal consistency.
  • 3D Variant: A more compute-heavy option using a fully 3D-convolution backbone (3D ResBlocks + TemporalConv + trilinear interpolation) for stronger temporal coherence.

Both variants support a scale factor between 1.0x and 4.0x. The 3D node further allows users to specify output size via a multiplier, target dimensions, or total megapixels.

Who it’s for

Users of ComfyUI who are generating videos with Minimax H3 and want to increase their output resolution faster and without the same quality loss associated with naive interpolation.

Highlights

  • Learned Upscaling: Uses a neural network trained on ~80,000 paired samples to produce sharper results than standard interpolation.
  • Two Backbone Options: Choice between a fast 2D model or a temporally-coherent 3D model.
  • Flexible Sizing: The 3D node supports scaling by multiplier, target dimensions, or megapixels with automatic pixel-grid alignment.
  • Performance Optimizations: Includes temporal chunking to manage VRAM for long videos and support for multiple precision levels (fp32, fp16, bf16) and devices (CUDA, ROCm, CPU).
  • Plug-and-Play: Integrates directly into ComfyUI as a custom node.

Related

  • Project
    AIMixer/ComfyUI_MiniMaxH3_Director

    A ComfyUI plugin that provides a director's console for MiniMax H3, enabling multi-segment video planning, synchronized audio generation, and advanced refinement tools for long-form AI video.

  • Project
    Rimagination/h3lite

    A local deployment skill for MiniMax H3 that enables AI agents to generate high-quality videos with native audio on Windows/NVIDIA hardware, optimized for low VRAM.

  • Dispatch
    MiniMax H3 Support in ComfyUI

    ComfyUI introduces day-zero support for MiniMax H3, an open-weights omni-modal video model capable of generating 2K video with native stereo audio on consumer hardware.

  • Project
    nkxx188/ComfyUI-MiniMaxH3-Easy

    A compact ComfyUI interface for MiniMax H3 that simplifies text-to-video, image-to-video, and reference-based video generation with a unified media input system.

  • Project
    huangserva/ComfyUI_MiniMaxH3_Director

    A collection of ComfyUI workflows for MiniMax H3 Director, enabling text-to-video, image-to-video, and advanced video-to-video editing with reference materials.