MiniMax H3 Conditioning

MiniMax H3 is the omni-modal video model that landed open-weights in August 2026: 33B, native stereo audio, 4–15 second clips at up to 2K/24fps, day-zero ComfyUI support. Great - and it generates short clips, so the moment you want anything longer you're into segment extension, and into hand-managing overlaps, lengths and prompts across every segment. That's what this node is for.

MiniMax H3 Conditioning takes your video as rows. Each row is one segment: a prompt, a duration, how much of the previous segment it continues from, and optionally a continuity override. Fill row 1, row 2 appears. Up to 24 of them. The node encodes every prompt up front, before any sampling starts, so your text encoder is loaded once instead of once per pass - which on a 33B pipeline is not a small favour.

The inputs that matter

clip wants a MiniMax CLIP - that's Load CLIP set to type minimax - and vae is the H3 video VAE, the same one used for frames and references. Reference audio needs audio_vae as well.

mode picks the task. t2va is prompt-only. i2va opens on first_frame. fl2va opens on first_frame and closes on last_frame. fl2va_batched builds every segment between a neighbouring pair of images you wire into images. ref2va builds every segment on the reference inputs, and those are the ones you address in the prompt by name - <Picture 1>, <Video 1>, <Audio 1> - with up to nine reference pictures, three reference clips (each with its own audio input) and three reference sounds. That's the mode for keeping the same people, places and voices across a whole run, and it's the one that needs ref2va plus at least one wired reference, or it raises.

aspect_ratio and megapixels set the canvas, with width/height overriding if you set either above 0 - the tooltip's own examples: 0.4 MP for 832×480, 1.0 MP for 1344×736. custom takes the shape from the first picture the mode reads, and 16:9 when there's nothing to read.

prompt_header and prompt_footer are the good bit of design. Header text goes before every segment's prompt and footer after, so the wardrobe notes, the subject definitions, the camera style and the audio bed live in one place instead of being pasted into twenty boxes. A blank line separates the header from the row's own prompt.

Then the rows: prompt_1, duration_1 (seconds, snapped to the model's frame grid), overlap_1 frames - 22 to carry on, 0 to cut somewhere new, 39 or 56 to hold the scene harder, ignored on segment 1 - and the optional continuity_1 override for what H3 Extend Window would otherwise do (carry, refresh, handoff, reference, cut, or as set). A blank prompt_1 ends the run, so you don't have to clean up trailing rows.

Outputs: prompts (the WAS_H3_PROMPTS bundle that drives the extend and clip-select nodes), segments (how many rows are live - wire this to a While Loop's count), latent and latents, and a report string stating the frames each segment came to after snapping.

Installing it

It's in WAS Node Suite v3 (MIT, WASasquatch), so install the pack and not the node:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/was-node-suite-comfyui.git

ComfyUI Manager, search WAS Node Suite v3, is the recommended route. ComfyUI 0.14.0+ and Python 3.10+; nothing is pip-installed and nothing is downloaded by the pack - and since H3's own weights are your download, features.network staying false in <ComfyUI user dir>/was-node-suite/config.yaml is doing what it says: a node missing weights will name the key it wants instead of quietly fetching gigabytes.

Before you build anything on H3

The weights are open but the licence is geofenced. The MiniMax H3 Community License grants broad rights below a $20M revenue line but scopes the whole grant to an Applicable Territory that excludes the EU, UK, South Korea and the United States - and the exclusion reaches outputs, not just the weights. The hosted Hailuo API is globally available; the local-weights path is not. Check where you stand before you invest a weekend in a pipeline.

Where it gets fiddly

Frame counts are snapped onto H3's 17k+5 grid - 124 frames is a real number, 120 is not - so the durations you type come back slightly different, and the report is how you find out. Overlap values shift on the same grid.

Everything downstream assumes this bundle. Sampling a row means feeding prompts and an index to MiniMax H3 Clip Select or H3 Extend Window, not poking at the conditionings directly. And a row's overlap only means something when something upstream is building continuity - the node's job is to describe the run, not to perform it.