MiniMax H3 Reference to Video

The sibling MiniMax H3 Image to Video node is text in, video out. This one is the grown-up version: same model, but you can hand it up to nine reference images, three reference videos, and audio - and then talk about them in your prompt by tag. Want a character to stay the same person across shots? Want a clip to borrow the motion and mood of footage you shot? Want the generated soundtrack to sound like a specific voice? That's what MiniMax H3's reference-to-video (ref2va) mode is for.

The tag system

The trick that makes this work is that every reference is addressable from the prompt. The node numbers them per type, 1-based, and you refer to them with angle brackets:

• <Picture 1>, <Picture 2>… - your images
• <Video 1>, <Video 2>… - your reference clips
• <Audio 1>, <Audio 2>… - your audio, including each reference video's soundtrack

They're presented to the model in a fixed order: images first, then videos (with each video's soundtrack labeled just before its video), then standalone audio. A prompt like "make <Picture 1> do the same dance as <Video 1>, with <Audio 1> as the backing track" is the shape you're aiming for. An unmentioned reference still conditions the output, just weakly - writing the tags is how you actually steer.

The inputs

Same base as the sibling (clip, vae, and now a required audio_vae, plus prompt/width/height/length), then:

• ref_image_size - match scales each reference down (keeping aspect) to your generation's pixel area; max keeps references at the pipeline's 2048px short edge for better identity fidelity. This is the one real trade knob: reference tokens ride through every sampling step, so max can be several times slower. Start with match.
• ref_images - up to 9; downscaled to a 2048px short edge if larger, never upscaled.
• ref_videos - up to 3, expecting 2–15s at 24 fps.
• ref_video_audios - the soundtrack paired to the same-numbered reference video.
• ref_audios - up to 3 standalone audio clips.

How it works

Every reference gets resized, VAE-encoded, and attached to the conditioning as a latent block that's re-injected on every sampling step - never denoised, just consulted. Videos are additionally shown to the Qwen tokenizer at 2 fps with timestamps so the model reads them as temporal, and audio goes through its own VAE (that's the audio_vae input). All of it rides the conditioning through the sampler, which is exactly why identity holds across a whole generated clip.

Gotchas

• Reference videos need at least 5 frames (~0.2s at 24 fps) or the node raises an error.
• Ref videos longer than your generation get truncated down to fit, snapped to the model's valid frame lengths - so feed roughly what you want back.
• The length knob still snaps to the 17k+5 grid, same as the sibling node.

It's brand new (August 2026), so treat your first runs as a test drive - tag one reference at a time until you've got a feel for how much it obeys. Once you do, this is the node that makes H3 feel like a studio tool instead of a text box: reference the face, reference the take, reference the voice, and it all lands in one generated clip with sound.