SOLRICKS/MiniMax-H3-Upscaler

MiniMax H3 Upscaler

Model files and example workflows for a custom ComfyUI node that upscales videos and enhances detail.

Upscale an existing H3 video with a learned latent upscaler, then refine detail with an H3 LoRA. The node uses ComfyUI's built-in model implementations and returns images and audio.

Comparison

Before — 1376×768	After — MiniMax H3 Upscaler · 1952×1088
Open original video	Open upscaled video

AI-generated source and upscaled output. Both clips: 243 frames, 24 FPS, 10.125 seconds. Video and audio streams are preserved without re-encoding.

Downloads

File	Location
h3_upscale_refiner_int8.safetensors	ComfyUI/models/loras/
minimax_h3_latent_upscaler_3d_conv_v1_fp16.safetensors	ComfyUI/models/latent_upscale_models/
MiniMax_H3_Upscaler.json	Load in ComfyUI
MiniMax_H3_Upscaler_API.json	API client

You also need a MiniMax H3 FL2VA diffusion model and H3 video VAE, such as the files in Kijai/MiniMax-H3-experimental. They are separate downloads and are not bundled here. The example workflow uses VideoHelperSuite for MP4 export.

Use in ComfyUI

1. Double-click an empty area of the ComfyUI canvas, type solricks in the node search, and select MiniMax H3 Upscaler.
2. Download the two model files into the folders listed above.
3. Load MiniMax_H3_Upscaler.json, choose your source video, and select the installed diffusion model and VAE filenames.
4. In MiniMax H3 Upscaler, select the FP16 latent upscaler and h3_upscale_refiner_int8.safetensors.
5. Start at 2.0 megapixels with temporal chunking and spatial tiling enabled. The target must be at least as large as the input; dimensions snap to H3's 32-pixel grid.

Use a diffusion model without generation/Turbo/style LoRAs. The upscaler node adds the refiner itself. The workflow preserves source FPS and passes source audio through. Its optional text/audio conditioning ports are left unconnected, so no text encoder or audio VAE is required by this example.

Licenses

The files have different licenses; see LICENSES.md and NOTICE. The INT8 refiner follows the MiniMax H3 Community License. The FP16 latent upscaler is Apache-2.0. SOLRICKS' node code license does not replace the model licenses.

This repository belongs in the Visual AI Models & Tools collection.