LBH-123-AI/Minimax_h3_latent_Upscaler

Minimax H3 Latent Upscaler

Neural latent-space upscaler for Minimax H3 video generation. Works directly on Minimax H3's 24-channel VAE latents to upscale spatial resolution (H×W) while preserving the time dimension.

Purpose

This model is designed to accelerate high-resolution H3 video generation:

1. Generate video at low resolution — far fewer latent tokens, much faster.
2. Upscale the latent in-place with this learned upscaler.
3. Re-sample / refine at the target resolution to recover detail.

By skipping the expensive decode → pixel upscale → encode round-trip through Minimax H3's heavy ~5B-parameter VAE, this pipeline saves a significant amount of generation time. It also avoids the ghosting / double-image artifacts that naive latent interpolation (bilinear/bicubic) introduces.

📸 Examples

Video upscale comparison — click to play:

(If the player doesn't render, download the video here.)

Image upscale comparison:

Files in this repository

File	Format	Precision	Size	Notes
minimax_h3_latent_upscaler_3d_bf16.safetensors	SafeTensors	bfloat16	~691 MB	Fastest on Ampere/Ada GPUs
minimax_h3_latent_upscaler_3d_fp16.safetensors	SafeTensors	float16	~691 MB	Best speed / memory balance
minimax_h3_latent_upscaler_3d_fp32.pth	PyTorch	float32	~1.38 GB	Highest numerical accuracy

All three checkpoints share the same 3D-convolution architecture. Pick the precision that matches your GPU and workflow.

Usage

Use with the companion ComfyUI custom node:

LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler

1. Install the custom node into ComfyUI/custom_nodes/.
2. Place the downloaded checkpoint in:

   ComfyUI/models/latent_upscale_models/
   

3. Add either "Minimax H3 Latent Upscaler (2D)" or "Minimax H3 Latent Upscaler (3D)" from the video/MinimaxH3 menu, connect a LATENT, pick the model, and set scale.

Supported upscale factors: 1.0× – 4.0× (continuous, 0.1 step; default 2.0×).

Training Data

The model was trained on ~80,000 paired samples (low-resolution latent + high-resolution target), balanced across modalities and scale factors to maximize generalization.

• ~70,000 video pairs
• ~8,000 2K image pairs

Scale distribution (approximate):

Scale	Share	Purpose
2×	40%	Dominant real-world factor
1.5×	10%	—
2.5×	10%	—
3×	10%	—
4×	10%	—
1.0×–4.0× (arbitrary decimals)	10%	Generalization to any in-between scale

Architecture

3D-convolution backbone with temporal convolution and trilinear interpolation. The architecture draws on and references the LTX 2.3 Spatial Upscaler (ltx-2.3-spatial-upscaler-x2-1.1.safetensors) and the neural-latent-upscaling approach pioneered by Ttl / ComfyUi_NNLatentUpscale.

License

Apache-2.0