MiniMax H3 Image to Video

MiniMax H3 is the newest heavyweight in ComfyUI core - an open-weight, multimodal video model from the Hailuo people, added in August 2026. Most local video models give you silent frames and you bolt audio on afterwards. H3 is different: it's built to generate picture and soundtrack from one denoising pass. This node is its front door. You type a prompt, optionally hand it a first and/or last frame, and it hands you the two things your sampler needs - a positive CONDITIONING and a LATENT - with the audio stream already packed inside the latent.

What it actually does

Think of it as the conditioning stage of a text-to-video / image-to-video pipeline. It does three jobs in one pass:

1. It tokenizes your prompt through the H3 CLIP (a Qwen3-VL-32B text encoder) into conditioning.
2. It builds the empty "AV latent" - a video stream plus an audio stream in a single nested tensor. That's the LATENT output, and it's why there's no separate audio stage in your graph: the audio is already in there.
3. If you fed it keyframes, it VAE-encodes them and pins them onto the conditioning as guides that get re-injected at every sampling step, never denoised.

The video side works like you'd expect: a diffusion transformer denoises frames. The audio side runs its own shifted schedule under the hood - you don't manage it, the model does.

The inputs that matter

Most of the schema is boilerplate, but two fields deserve attention:

• length - Frame count at 24 fps, snapped up to the model's 17k+5 grid. 124 is about 5 seconds. The trained range is roughly 124–362 frames; beyond that is untested and quality wanders. Ask for 120 and you'll get the next valid grid number back, so don't be surprised when the output runs a frame or two longer than you asked.
• first_frame / last_frame - Optional images. The first frame is stretched to the canvas (the "geometry anchor"), the last frame is cover-cropped (the "follower"). Feed both and you get a video that starts on one image and lands on another; feed just first_frame and you have classic image-to-video.

width/height default to 1344×768 and must be multiples of 32.

Wiring it up

Both outputs feed the sampler: positive → KSampler's positive input, LATENT → its latent input. The node itself loads nothing - it consumes the clip and vae from wherever you loaded the H3 checkpoint (the model lives on HuggingFace as MiniMaxAI/MiniMax-H3; load it with Comfy's normal loaders and the structure is auto-detected). The flow-shift defaults for the optional ModelSamplingMiniMaxH3 patch (12/3) are already baked into the checkpoint, so you don't need it unless you want to tune.

Where people get burned

• Length surprises. Not a bug - it's the frame grid. Read the tooltip, expect snapping.
• New-model rough edges. This shipped weeks ago as of writing. If a sampler misbehaves, try another - H3's audio schedule made the early samplers unhappy, and the fix history in the repo is real.
• Don't look for reference identity here. For character consistency, the sibling MiniMax H3 Reference to Video node is the one with reference images, clips and audio. This node is plain text-plus-keyframes.

The bottom line: if you want to try the most interesting open video model of the summer, this is the node that gets your prompt into it - video and audio, one pass, one graph.