MiniMax H3 VideoBatch

Some H3 shots aren't anchored by a single still - they're driven by an existing clip you want to extend, restyle, or blend with another reference. That's the gap this node fills. It takes up to three independent video frame-batches and packs them into a single MINIMAX_H3_VIDEO_BATCH object, with each video labeled (<Video 1>, <Video 2>, <Video 3>) so a prompt can point at a specific reference instead of a blob of frames.

How it works

It's deliberately dumb and that's the point. Each video_N input is a [T, H, W, C] IMAGE tensor - the standard way ComfyUI represents a frame batch. The node validates the shape (4D, 3 or 4 channels), strips alpha down to RGB, computes the clip's duration at 24 fps (frames / 24), and tags it with its label. Output is a list of those descriptors, ready for whichever node consumes a video reference batch.

There's no resizing, no interpolation, no magic - if you want the references preprocessed, you do that upstream. It's a packing node, and the value is mostly in the labeled structure: downstream prompt tooling can say "extend <Video 2>" and mean a specific clip.

The inputs and outputs

• video_1, video_2, video_3 - all optional IMAGE inputs. You need at least one; the node throws a clear "at least one video input is required" error otherwise. There's a hard cap of three, baked into the source as MAX_VIDEOS = 3.

One output: VideoBatch, type MINIMAX_H3_VIDEO_BATCH, which plugs into reference-driven generation (for example feeding into the Conditioning pipeline for full-reference work, alongside the PackageData path for mixed media).

Installing it

The pack-wide routine - ComfyUI Manager search "MiniMax H3", or:

cd ComfyUI/custom_nodes
git clone https://github.com/xiaolibai-sys/ComfyUI-MiniMaxH3
pip install -r requirements.txt

then restart. No models needed - pure data plumbing.

Common issues

• "At least one video input is required." Obvious but worth stating: all three sockets are optional, and an empty node is an error. Wire at least one.
• Shape errors. Each input must be [T,H,W,C] with 3 or 4 channels. If you feed it a single image ([H,W,C]), that's the error, exactly as the message says.
• Wrong duration than expected. Duration is computed at 24 fps from frame count - if your loader produced 30 fps footage, the batch's duration math won't match real time. Feed frame batches at a consistent fps.