Empty MiniMax H3 AV Latent (Batch) ⚡

MiniMax H3 is a 33B omni-modal video model: text, image, video and audio all live in one context, and the clip comes out with native stereo audio. But an omni-modal model needs an omni-modal starting point - a latent that holds the video stream and the audio stream at once. This node builds that, in H3's specific frame grid, with a batch size on top.

Why the shape matters

Like the LTX-2.5 empty-AV node in the same pack, this is a geometry node: it creates the noise container, and getting the geometry wrong means the sampler chokes. But H3's grid is its own thing. Duration snaps to the model's 17k+5 frame grid at 24 fps - align_frame_count() in the source just bumps the frame count up until n % 17 == 5. The default length of 124 lands at ~5 seconds; the trained range runs roughly 124–362 (about 15 seconds), and the tooltip is refreshingly honest that anything longer is untested.

So the defaults you see are there for a reason: 1344×768, 124 frames, batch 1. Those are the model's comfort zone, and stepping outside it is where quality - or the sampler - starts to complain.

The inputs that matter

• width / height - spatial size in steps of 32 (defaults 1344×768, a 16:9-ish frame).
• length - frame count at 24 fps, auto-snapped up to the 17k+5 grid. Want ~10 seconds? 243 gives you ~10.13s.
• batch_size - videos per job, each with distinct noise from the sampler seed. This is the star feature and it comes with a condition: for batch_size above 1, route the model through MiniMax H3 Batch Patch. H3's DiT only accepts batch 1, so the patch node isn't optional - it's what makes the batch actually sample.

One output

LATENT - the joint video+audio latent, straight into your H3 sampler.

How it fits the wider workflow

MiniMax H3's prompt side is unusually strict, and this pack gives you the matching tooling: MiniMax H3 Prompt Writer turns a raw idea into a schema-correct H3 prompt (or MiniMax H3 Prompt Format wraps fields you wrote by hand). Those produce the exact envelope the model expects, and this empty latent is the other end of that pipeline - prompt in, shaped noise, clip out.

Installing it

Part of the ComfyUI-GGUF-Loader pack under 🤖 CCTech/MiniMax H3. ComfyUI Manager → search "ComfyUI-GGUF-Loader" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ChrisColeTech/ComfyUI-GGUF-Loader
cd ComfyUI-GGUF-Loader
pip install -r requirements.txt

Common issues

Two classics. First: set batch_size to 3, forget the Batch Patch node, and the sampler dies on a batch-axis error - the patch is required, not recommended. Second: grab the LTX-2.5 empty-latent node instead, because you saw "AV latent" and reached for the first one, and get a channel mismatch. H3 and LTX-2.5 share the nested-latent idea but not the dimensions; use the node that matches the model you loaded. If in doubt, check which model your loader actually gave you - that tells you which empty node you need.