MiniMax H3 To Video

The one-stop H3 conditioning node

MiniMax H3 is the 33B omni-modal video model that landed open-weights in August 2026 and immediately took the top spot for video editing with audio - it treats text, images, video and audio as one input context and generates clips with native synced stereo sound. The catch for ComfyUI users: H3's conditioning is split across two different core nodes. MiniMaxH3ImageToVideo handles text-to-video and first/last-frame modes; MiniMaxH3ReferenceToVideo handles the reference (r2v / rv2v) side. Remembering which one to reach for, and how to wire a pile of references into it, is exactly the kind of friction this pack exists to remove.

easy minimaxH3ToVideo is the single entry point. Pick a mode, connect whatever media you have, and it routes to the right core node for you - including splitting videos into their frames-plus-audio components so the references land where H3 expects them.

How it works

Under the hood it's a graph-building node. It reads your images, audios, and videos inputs (all lists get expanded automatically - so yes, an IMAGE batch works), then picks a path:

• No media, or mode = multi_frames / last_frame → routes to MiniMaxH3ImageToVideo, feeding first_frame and/or last_frame from your image list.
• mode = reference with media → routes through the pack's own reference bridge (see the sibling node, Easy MiniMax H3 Reference Bridge), which maps your images to ref_image_0…, videos to ref_video_0… plus their extracted audio, and standalone audio to ref_audio_0….

There are hard caps: 9 reference images, 3 reference videos, 3 reference audios. Try to sneak in a fourth video and it raises instead of silently dropping it.

The inputs that matter

• mode - reference (default), multi_frames, or last_frame. This is the big dropdown that changes everything about how your media is used.
• prompt - the text prompt, multiline and dynamic-prompt-aware.
• clip, vae, audio_vae - audio_vae is optional unless you're passing reference audio, in which case it's required (the node tells you, loudly).
• length - frame count at 24 fps, default 124 (≈5s). The tooltip says it all: it snaps up to the model's 17k+5 grid and the trained range is roughly 124–362 - longer is untested. Don't expect magic at 900 frames.
• ref_image_size - match (default) scales each reference down to the generation's pixel area; max uses the reference pipeline's 2048px short edge for the best identity fidelity. Real trade-off here: reference tokens ride through every sampling step, so max can be several times slower. If faces drift, try max; if you're just iterating, stay on match.

Outputs are positive (CONDITIONING) and latent (LATENT) - straight into your sampler, with the model's audio VAE handling the sound side.

Setup and the elephant in the license

Install is the standard pack flow (Manager → ComfyUI-Easy-Media, or git clone into custom_nodes + restart). You'll want a reasonably current ComfyUI since it ships the H3 core nodes - the pack includes fallback implementations for older builds, but the real thing is better.

One thing nobody should skip, because it's a legal gotcha rather than a technical one: the MiniMax H3 Community License excludes the US, EU, UK, and South Korea from its Applicable Territory. If you're in one of those regions, you're not licensed to run the local weights at all - the hosted Hailuo API is a different matter. The node works fine technically; the license is your call to make before you click queue.

For everyone else, this is the node that makes H3 feel like one model instead of three. Wire media in, pick a mode, sample.