PickMinimax_h3_latent_Upscaler: Minimax H3 VAE Bypass for 4x Resolution Video Generation

Significantly reduces the time required for high-resolution rendering when generating videos based on Minimax H3. It skips the decode-pixel upscale-encode process that goes through a VAE with approximately 5 billion parameters, instead directly increasing spatial resolution from the 24-channel VAE latent. It expands only the H×W resolution while keeping the temporal dimension unchanged.

It prevents ghosting and double-image artifacts that occur when upscaling latents using simple interpolation methods like bilinear or bicubic. The model is trained with a structure combining a 3D convolution backbone, temporal convolution, and trilinear interpolation. It was designed with reference to the LTX 2.3 Spatial Upscaler and Ttl's neural-latent-upscaling approach.

It was trained on approximately 80,000 low-resolution to high-resolution paired samples, including 70,000 videos and 8,000 2K images. Generalization performance was improved by reflecting diverse scale distributions, with 2x upscaling accounting for 40%. It supports continuous upscaling factors from 1.0x to 4.0x in increments of 0.1.

It is designed to be used with ComfyUI custom nodes. It provides checkpoints in three precisions—bfloat16, float16, and float32—allowing selection based on GPU specifications. Distributed under the Apache-2.0 license, it is suitable for developers looking to resolve bottlenecks in high-resolution video generation pipelines.