LBH-123-AI/Minimax_h3_latent_Upscaler · Hugging Face

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

https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler#minimax-h3-latent-upscalerMinimax H3 Latent Upscaler

Neural latent-space upscaler forMinimax H3video generation. Works directly on Minimax H3’s24-channel VAE latentsto upscale spatial resolution (H×W) while preserving the time dimension.

https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler#purposePurpose

This model is designed toaccelerate 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 expensivedecode → pixel upscale → encoderound-trip through Minimax H3’s heavy ~5B-parameter VAE, this pipeline saves a significant amount of generation time. It also avoids theghosting / double-image artifactsthat naive latent interpolation (bilinear/bicubic) introduces.

https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler#%F0%9F%93%B8-examples📸 Examples

Video upscale comparison— click to play:

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

Image upscale comparison:

https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler#files-in-this-repositoryFiles in this repository

FileFormatPrecisionSizeNotesminimax\_h3\_latent\_upscaler\_3d\_bf16\.safetensorsSafeTensorsbfloat16~691 MBFastest on Ampere/Ada GPUsminimax\_h3\_latent\_upscaler\_3d\_fp16\.safetensorsSafeTensorsfloat16~691 MBBest speed / memory balanceminimax\_h3\_latent\_upscaler\_3d\_fp32\.pthPyTorchfloat32~1.38 GBHighest numerical accuracy All three checkpoints share the same3D-convolution architecture. Pick the precision that matches your GPU and workflow.

https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler#usageUsage

Use with the companion ComfyUI custom node:

LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler

1. Install the custom node intoComfyUI/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 thevideo/MinimaxH3menu, connect aLATENT, pick the model, and setscale.

Supported upscale factors:1.0× – 4.0×(continuous, 0.1 step; default2.0×).

https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler#training-dataTraining 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.

• ~70,000video pairs
• ~8,0002K image pairs

Scale distribution (approximate):

ScaleSharePurpose2×40%Dominant real-world factor1.5×10%—2.5×10%—3×10%—4×10%—1.0×–4.0× (arbitrary decimals)10%Generalization to any in-between scale

https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler#architectureArchitecture

3D-convolution backbone with temporal convolution and trilinear interpolation. The architecture draws on and references theLTX 2.3 Spatial Upscaler(ltx\-2\.3\-spatial\-upscaler\-x2\-1\.1\.safetensors) and the neural-latent-upscaling approach pioneered byTtl / ComfyUi_NNLatentUpscale.

https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler#licenseLicense

Apache-2.0