Innovate Futures @ Benji

MiniMax H3 Latent Upscale in ComfyUI - Multi-Stage Sampling (Patreon)

Published:

2026-08-25 14:58:52

Edited:

2026-08-30 08:50:29

Imported:

2026-09

Content

Video : https://youtu.be/eUhIsMHfmsA

New video is up. MiniMax H3 finally got latent upscale models on Hugging Face, and this one shows you how to run them in ComfyUI as a second-stage pass. Generate at 0.6 or 0.7 megapixels, then the H3 model itself re-samples the latent up to 1080p. Real detail, not just bigger pixels.

I also show the fork I made of the MMH3 Ultimate Upscale node. The original loads and offloads the upscaler on every single chunk, which is pure wasted time on a 24GB+ card. My version keeps the model resident, and I demo the difference.

Demos include a text-to-video fight scene and reference-to-video characters, upscaled from 480p all the way to 1080p. The full workflow is attached right here, turbo LoRA included.

LATENT UPSCALE MODEL (Hugging Face):

Repo: LBH-123-AI/Minimax_h3_latent_Upscaler https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler/

Examples: https://github.com/benjiyaya/ComfyUI-MinimaxH3-Latent-Upres/blob/main/Minimax_h3-Turbo_t2vi2v-LatentUpres.json

CUSTOM NODE (GitHub):

Repo: https://github.com/benjiyaya/ComfyUI-MinimaxH3-Latent-Upres

Base upstream: bbaudio-2025/Comfyui-MMH3-UltimateUpscale

Thanks for supporting the channel.

— Benji

Files

MiniMax H3 Latent Upscale in ComfyUI - Multi-Stage Sampling

In this video I walk through the new latent upscale models for MiniMax H3 on Hugging Face and the ComfyUI-MinimaxH3-Latent-Upres custom node that runs them. I show the full second-stage upscaling workflow, similar to LTX 2.5's second stage: generate at a low megapixel setting like 0.6 or 0.7, then upscale up to 2 megapixels (1080p) using the H3 model itself. I explain why I added a keep-models-resident option so high-VRAM cards stop wasting time loading and offloading the 600MB upscaler on every chunk, and I demo reference-to-video (Ref2VA) generations upscaled from 480p to 1080p, including Tekken and Mortal Kombat style fight scenes drafted with an AI agent. ComfyUI users generating AI video with MiniMax H3 who want sharper, higher-resolution results without leaving latent space. If you run H3 text-to-video or reference-to-video and want a second-stage detail pass, or if you have a 24GB+ GPU and noticed slow load/offload cycles in the original node, this is for you. Upscaling happens in latent space before VAE decode, so the H3 model re-samples every tile with the original conditioning instead of just sharpening pixels. Audio stays intact because the node understands H3's nested video+audio latent. That means real detail reconstruction on long videos, on any VRAM, with one node. LATENT UPSCALE MODEL (Hugging Face): Repo: LBH-123-AI/Minimax_h3_latent_Upscaler https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler/ Examples: https://github.com/benjiyaya/ComfyUI-MinimaxH3-Latent-Upres/blob/main/Minimax_h3-Turbo_t2vi2v-LatentUpres.json Blog Post : https://www.patreon.com/aifuturetech/posts/minimax-h3-in-167653244 CUSTOM NODE (GitHub): Repo: https://github.com/benjiyaya/ComfyUI-MinimaxH3-Latent-Upres Base upstream: bbaudio-2025/Comfyui-MMH3-UltimateUpscale Calliope - Local-first AI Idea-to-video studio https://github.com/benjiyaya/Calliope Local Workstation GPU : https://amzn.to/3XfXsAO -------------------------------------------------------------------------------------------------------------------------------- If You Like tutorial like this, You Can Support Our Work In Patreon: https://www.patreon.com/c/aifuturetech