LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler

What it solves

This project provides a way to upscale video latents for the Minimax H3 model without the slow and computationally expensive process of decoding latents to pixels, upscaling them, and then re-encoding them back into latents. By performing the upscale directly in the latent space, it significantly accelerates high-resolution video generation and avoids common artifacts like ghosting or double-images caused by basic interpolation.

How it works

It introduces custom ComfyUI nodes that use trained neural networks to upscale 24-channel Minimax H3 VAE latents. It offers two primary neural upscaler variants:

• 2D Upscaler: A lightweight ResBlock backbone with temporal 3D-Conv layers for spatial upscaling while maintaining temporal consistency.
• 3D Upscaler: A fully 3D-convolution backbone that processes the spatiotemporal volume jointly for stronger coherence, supporting scaling by multiplier, target dimensions, or total megapixels.

Additionally, it includes the MMH3 Split Upscale combo, which is a tiled, "hires-fix" style re-sampling upscaler. Instead of a pre-trained network, it splits the AV latent into temporal chunks and spatial tiles, re-runs the diffusion sampler on each piece at a higher resolution, and stitches them back together using seam-denoise, two-level color matching, and temporal anchors to prevent flickering and seams.

Who it’s for

Users of ComfyUI and the Minimax H3 video generation model who want to generate high-resolution videos faster and with better visual quality and temporal stability.

Highlights

• Neural Latent Upscaling: Uses trained networks instead of naive interpolation for sharper results.
• Two Backbone Options: Choice between a fast 2D variant and a more coherent 3D variant.
• Tiled Re-sampling: A decomposed and optimized tiling system (MMH3 Split Upscale) to handle high-res videos within VRAM limits.
• Advanced Stitching: Features like seam-denoise, color matching, and triple temporal anchors to eliminate ghosting and inter-tile flicker.
• Flexible Sizing: Supports scaling by multiplier, specific target dimensions, or megapixels with automatic pixel-grid alignment.