MiniMax H3 Day-0 Support in ComfyUI: Open Weights, Native Audio, and 2K Video

Overview

MiniMax H3 is an open-weights omni-modal video model that accepts text, image, video, or audio inputs and generates video with real stereo sound, up to 2K resolution and up to 15 seconds per clip.

Model Capabilities

  • Text-to-video: generate video from a prompt alone.
  • Image-to-video: animate a supplied image.
  • First-and-last-frame: control the opening frame, the closing frame, or both and let the model fill the intermediate frames.
  • Reference-to-video: supply reference images, video, or audio to carry a subject, motion, or voice through the generated clip.
  • Multimodal context understanding: the model jointly processes images, audio, and video together with a prompt that describes how they relate, collapsing several separate tasks into one model.
  • Native stereo audio: audio is produced as a property of the model in the same pass, not added as a post‑process.
  • Editing and motion transfer: a reference video can provide motion (camera move, performance, cutting rhythm) while subject and style come from other inputs, enabling iterative shot editing.

Optimizations for Local Inference

The ComfyUI team reduced the memory footprint of MiniMax H3 by pruning the model’s modulation weights (about 40 % of total parameters) and replacing them with a functionally equivalent lookup table, which incurs no loss in output quality. The weights also ship with an accurate int8 convrot quantization and custom kernels that lower peak VRAM use. As a result, the total memory footprint drops from 123.6 GB in full precision to 42.5 GB for the smallest variant. Combined with dynamic VRAM offloading, this allows the model to run locally on a GPU such as the RTX 3060.

Getting Started in ComfyUI

  1. Update ComfyUI to version 0.30.0 or later, or use Comfy Cloud.
  2. Download one of the provided workflows from the template library:
    • MiniMax H3 I2V workflow
    • MiniMax H3 R2V workflow
    • MiniMax H3 T2V workflow
  3. Follow the workflow note to download the model weights from Hugging Face (Comfy‑Org/MiniMax‑H3) and place them in the correct model directory.
  4. Write your prompt, connect any frame or reference inputs, and run the workflow.

Community Feedback and Discussion

  • A user with an RTX 4070 Ti Super (16 GB VRAM) reported that generating a 10‑second 480p video takes about 10 minutes, describing the results as spectacular.
  • Another user asked how long a 15‑second clip would take on a 16 GB RTX 3060, indicating interest in inference time on the target hardware.
  • One comment praised the mouse render in the example outputs, noting several clips as a clear step forward compared to current SOTA models, while pointing out a remaining "AI smoothening" artifact in a beverage‑ad clip.
  • A different commenter found the aesthetic of the samples "painfully bland and generic."
  • A user highlighted the reference‑to‑video mode as the missing piece for independent cinematography, enabling proper stitching of different scenes without extensive rework.
  • Another user said the samples made previous models (LTX2 and WAN) feel "completely worthless" and expressed excitement about the quality.
  • A comment noted that human directors remain valuable because they can arrange AI‑generated shots into a cohesive narrative, similar to how an EDM producer arranges pre‑made elements.
  • A user inquired about running the model on Mac hardware (e.g., Mac Studio Ultra).
  • One user shared personal experiments, observing that while normal scenarios work well, weirder concepts tend to break down, but they were impressed by the speed and quality of text‑to‑video generation on an RTX GPU.

These community reactions highlight both the enthusiasm for MiniMax H3’s open‑weights release and native audio/video capabilities, and the practical considerations around performance, aesthetics, and workflow integration.

Sources