MiniMax H3 Context IR (Prompt Enhancer)

MiniMax H3 is one of the most capable closed video models around, and the thing people consistently report is that it rewards good prompts - the long, structured, shot-by-shot kind. Writing those is a skill, and every generation burns Comfy credits, so failing on the first try gets expensive. This node is the shortcut. You type a rough description, optionally hand over the images, video, or audio you want the final clip to respect, and it returns a polished, structured prompt ready to feed into any of the MiniMax H3 video nodes.

The name is scarier than the job. "Context IR" sounds like a research paper; from your side it's just a prompt enhancer that also looks at your media. Here's the mechanism: Comfy uploads whatever you attached to its servers, posts text plus media to the H3 Context IR endpoint, then polls the task until it resolves (roughly every five seconds). What comes back is one STRING - the enhanced prompt. That prompt doesn't merely describe the scene, it references your media by position: "Image 1", "Video 2", "Audio 3". Which is why the docs keep insisting you re-attach the same media, in the same order, to the generation node. Get the order wrong and the prompt will happily talk about the wrong subject.

The inputs that actually matter

Most of this lives inside the model combo - pick "MiniMax H3" and the rest unfolds:

• prompt - your raw description. The only required bit; keep it honest, the enhancer does the dressing up.
• duration - 4 to 15 seconds, matching the video you intend to generate.
• ratio - the aspect ratio the enhanced prompt is tuned for. adaptive is the default and is great, with one trap: it requires at least one piece of media attached. Text-only jobs need an explicit ratio like 16:9 or 9:16.
• reference_images / reference_videos / reference_audios - up to 9 images, 3 videos (2–15s each, 15s total), 3 audio clips. Audio can't ride alone; you need an image or video alongside it.

Separately, first_frame and last_frame exist for image-to-video-style jobs. They are mutually exclusive with reference media - pick one lane.

The output and where it goes

You get a single STRING. Wire it into the prompt input of MinimaxHailuo03TextToVideoNode, MinimaxHailuo03FirstLastFrameNode, or MinimaxHailuo03ReferenceNode - whichever matches the media you attached - and re-attach that same media there in the same order. The enhancement itself costs only a fraction of a generation, priced per call and billed to your Comfy account credits, which is why it's worth doing even for a short clip.

Getting it and the gotchas

It ships with ComfyUI core - no install, nothing to clone. It landed with the H3 node family in mid-2026, so you need a reasonably current ComfyUI. And because it's an API node, it needs you signed into your Comfy account with credits on it.

Where people get burned:

• adaptive with no media - the node raises an error. Pick an explicit ratio.
• Order mismatch - enhanced prompt says "Image 1", but you attach a different image first at the generation node. The output won't match what the enhancer planned.
• Media that doesn't pass validation - images need an aspect ratio between 0.4 and 2.5 and at least 256×256; videos must be 23.976–60 FPS and under 15s total. The node tells you exactly what's wrong, which beats a mid-generation failure.
• Empty response / silent failure - usually a credits or account problem on the Comfy side; check your balance before blaming the prompt.

One honest caveat: an enhanced prompt is a strong starting point, not a guarantee. H3 is good enough that it sometimes over-commits to what the enhanced text literally says, so read the returned prompt once before you spend the credits.