Minimax H3 Latent Upscaler

Minimax H3 Latent Upscaler

Upscales Minimax H3 video latents directly, so you can generate at low resolution and refine at the target size without a decode and re-encode. For people running Minimax H3 video generation.

Made by LBH-123-AI on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: A neural upscaler that works on Minimax H3's 24-channel VAE latents. It raises the spatial size, height and width, while leaving the time dimension alone. It speeds up high-resolution video generation, since you can generate small, upscale the latent, and refine at the final size.

Who it suits: People who generate video with Minimax H3 and want faster high-resolution output without a round trip through the heavy VAE.

Industry | Who uses it | How they use it | Example
Film and media | Video creators | Generate at low resolution, then upscale the latent | Draft a clip small and finish it at a larger size
Marketing | Content teams | Produce high-resolution clips faster | Render a product video with less waiting
Research | Generative video researchers | Compare latent upscaling with plain interpolation | Test 2D and 3D versions across scales

If you do not use Minimax H3, you can skip it, since it works on that model's latents.

Where to use it:

• Minimax H3 video generation pipelines
• Windows, macOS or Linux, for downloading the files
• Image generation as well, since the card shows an image upscale comparison

Use case | Who | What they get back
Latent upscale | Pipeline builders | A larger latent that keeps the time dimension
Faster high-resolution video | Creators | Less time than a decode, pixel upscale and encode round trip
Sharpness control | Power users | Six tiers with the 2D release, from most faithful to sharpest

How to set it up:

• Python with pip
• A Minimax H3 generation setup that produces and consumes its 24-channel latents
• The checkpoint you choose, 2D or 3D

1. Install the Hugging Face hub package.
2. Download the repo.
3. Add the upscaler between your low-resolution sampling pass and the final refine pass.

pip install -U huggingface_hub
hf download LBH-123-AI/Minimax_h3_latent_Upscaler

How to use it: Generate the video at low resolution, upscale the latent with this model, then resample at the target resolution to recover detail. With the 2D release you pick a sharpness tier from 1 to 6.

Tool | What it does
minimax_h3_latent_upscaler_2d_conv_v1 | The 2D release, with about half the parameters and six sharpness tiers
3D release | The earlier version, with no sharpness control

Example: You render a clip at a small size, pass its latent through the 2D upscaler at tier 1 for the most faithful result, then refine at the full size. The pass avoids the ghosting that simple interpolation adds.

For newcomers and experts: If you are new, start with the 2D release at tier 1 and a modest scale, then try higher tiers if you want a sharper look.

The 2D release has 182,140,312 parameters, 52.8 percent of the 3D one at 345,280,216. In the card's test on 2,500 validation pairs and 36 scales from 1.1 times to 4.0 times, it cut latent error by 84.7 percent against trilinear, against 61.1 percent for 3D. The 3D release gets noticeably worse below 1.5 times.

Generate at low resolution
Upscale the latent with the model
Resample at the target resolution
Decode the final video

Good to know:

• Trilinear sharpness reads above 1.0 because of ringing, not extra real detail
• The 3D release lacks sharpness tiers and degrades below 1.5 times
• License: Apache 2.0
• Cost to run: none named, you supply the hardware
• Data it reaches: the latents you supply and the files downloaded from Hugging Face
• Not for: use with models other than Minimax H3

Title: Decoded PSNR, mean over 36 scales
Trilinear: 23.54 dB
3D_Conv_v1: 29.30 dB
2D_Conv_v1: 30.47 dB

On Hugging Face, 13 November 2026

• Library: minimax-h3
• Formats: Safetensors
• Downloads in the last 30 days: 354,859
• Likes: 489
• License: apache-2.0
• Quantized versions: 1
• Fine-tunes: 5
• Tags: minimax-h3, latent_upscaler, latent-upscaler, latent-space-upscaling, video-super-resolution, super-resolution, comfyui, safetensors, pytorch, 3d-convolution, temporal-convolution, latent-diffusion

Links

• Model: https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler
• Files: https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download LBH-123-AI/Minimax_h3_latent_Upscaler

Then follow the usage in its model card. Nothing in this kit runs by itself.