MiniMax H3 Latent Upscaler 3D: Neural Upscaling With ComfyUI Nodes

What it is

The model works directly on MiniMax H3's 24-channel VAE latents to upscale the spatial resolution (H × W) while preserving the time dimension. A trained neural network replaces naive bilinear/bicubic latent interpolation, which avoids the ghosting and double-image artifacts naive upscaling introduces.

Its main purpose is to accelerate high-resolution H3 video generation:

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

By skipping the slow decode → pixel upscale → encode round trip through H3's heavy ~5B-parameter VAE, this pipeline saves a significant amount of generation time. Note that it saves time, not VRAM — the refinement pass still runs at the target resolution, so peak memory is comparable to generating high-res directly.

The custom node pack ships two node variants under the video/MinimaxH3 category:

• Minimax H3 Latent Upscaler (2D) — a 2D ResBlock backbone with temporal 3D-conv layers, lightweight and fast, using a simple scale factor (1.0×–4.0×).
• Minimax H3 Latent Upscaler (3D) — a fully 3D-convolution backbone that processes the spatiotemporal volume jointly for stronger temporal coherence, with three resize modes in one node: scale by multiplier, target dimensions, and megapixels (with pixel-grid alignment and aspect-ratio lock).

Both nodes support upscaling only (effective scale >= 1.0) and offer fp32 / fp16 / bf16 inference. Recent updates added an enable_chunking toggle for long videos (with temporal-chunk edge blending), automatic CPU offload after execution, and ROCm (AMD GPU) backend support.

Examples

The repository ships a video and an image upscale comparison:

Video upscale comparison (click to play)

Image upscale comparison

ComfyUI usage

1. Clone Comfyui_Minimax_h3_latent_Upscaler into ComfyUI/custom_nodes/ and restart ComfyUI. No extra Python packages are required.
2. Download one of the checkpoints (bf16 / fp16, ~691 MB each, or the fp32 .pth) into ComfyUI/models/latent_upscale_models/. The loader auto-detects the architecture, so one checkpoint works for both the 2D and 3D nodes.

Typical workflow: [Low-res H3 Latent] → [H3 Latent Upscaler] → [Refine / Re-sample] → [VAE Decode]. The node pack includes example I2V and R2V workflows:

The earlier H3 2x latent upscaler by Mamad8 targets clean post-sampling latents; this release takes a similar latent-space approach with a heavier 3D backbone that also supports direct target-dimension and megapixel resizing.