MiniMax H3 2x Latent Upscaler With ComfyUI Nodes

Mamad8 released a 2× latent upscaler for MiniMax H3 on August 8, together with a pair of minimal ComfyUI nodes (model card, custom nodes). The model doubles the spatial dimensions of a fully denoised H3 video latent, so a workflow can stay in latent space instead of paying for a VAE decode → pixel resize → VAE re-encode round trip.

What it is

This is not a conventional image or video upscaler. It does not make a finished render sharper, and a direct decode of its output can look softer than the original. Its purpose is to move a clean H3 video latent onto a 2× larger spatial latent grid quickly, so high-resolution continuation or a second sampling pass can run on a bigger canvas.

• Doubles only the spatial latent dimensions; temporal length is unchanged.
• When used through the companion node, the audio stream in H3's joint video/audio latent is preserved untouched.
• Accepts clean latents only — applying it to intermediate noisy latents is not supported.

How it was trained

Training pairs were built from clean H3 latents: each low-resolution latent was decoded with the H3 VAE, enlarged 2× in pixel space with Lanczos, then deterministically re-encoded to provide the teacher latent. The lightweight network learned a correction on top of bilinear latent interpolation using latent and decoder-aware reconstruction, SSIM, spatial-consistency and temporal-consistency losses. The H3 generator itself was not trained or modified.

ComfyUI usage

1. Clone ComfyUI-H3-Latent-Upscaler-Mamad8 into ComfyUI/custom_nodes/ and restart ComfyUI. No extra Python packages are required.
2. Download h3_clean_latent_upscaler_v1_mamad8.safetensors (~56 MB) into ComfyUI/models/h3_latent_upscalers/.

The node package provides two nodes: Load H3 Latent Upscaler and Upscale Clean H3 Latent 2x. Place the apply node after sampling finishes and before the normal VAE decode:

H3 sampler → Upscale Clean H3 Latent 2x → VAE Decode

For example, a clean latent produced for 672×384 decodes at 1344×768 after 2× latent upscaling. Frame count and audio are unchanged. Continuing H3 denoising after the upscale requires an explicit re-noising and high-resolution continuation workflow, which is intentionally left outside the two nodes.

The repository ships a default workflow that uses dependency-free core ComfyUI video and VAE nodes:

Community notes

Banodoco community members started testing it the same day. RuneX ran a two-pass setup — a full 25-step first pass at low resolution, a 2× latent upscale, then an 8-step second pass — and reported the audio came out noticeably richer (test notes). mamad8 noted the release partially preempts an official upscaler that MiniMax itself is expected to ship for H3.

For upscaling between samplers (re-noise + second pass inside one workflow), Tr1dae's ComfyUI-MiniMaxH3_LatentUpscaler takes a different approach with an audio_denoise control for how much of the pass-1 audio to re-mix.