Trying out the MiniMax H3 Motion-Context custom node in ComfyUI

This is a custom node designed to reference a few to several dozen frames from the end of an input video to generate a continuation that is as natural as possible.
It feels like a video extension technology.

I have explained MiniMax H3 several times before, so please take a look if you are interested.

Environment

OS: Windows 11
GPU: GeForce RTX 4090
CPU: i9-13900KF
Memory: 128GB

Setup Procedure

CUDA 13 is required, so please refer to these steps or similar resources to update your CUDA version.

Install the following custom node and update ComfyUI to the latest version.

cd ComfyUI\custom_nodes
git clone https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context.git
git clone https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3.git

Models

minimax_h3_fl2va_pruned_int8_convrot.safetensors download and place in "ComfyUI\models\diffusion_models"qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors download and place in "ComfyUI\models\text_encoders"
minimax_h3_audio_vae_fp32.safetensors download and place in "ComfyUI\models\vae"

minimax_h3_video_vae_int8_convrot.safetensors download and place in "ComfyUI\models\vae"

minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy_resized_avg_rank_21_bf16.safetensors download and place in "ComfyUI\models\loras"

Workflow

The node on the left generates a video that extends from the uploaded video.
The image is for I2V (Image-to-Video) to be used as the video generated for the extension.

Generation Results

The generation results are as follows.
The generation time was 48 seconds, CPU memory was 63GB, and GPU memory was 59GB (21GB VRAM + 38GB shared memory).

Impressions

The audio is not perfect.
It seems the singing voice is being reflected, so I feel that if I make the context a bit longer, it might work better.

With video extension like in LTX-2.3, you generally cannot specify the final frame for the subsequent video, so I feel this feature itself has some practical uses.