MiniMax H3 is now inside CapCut, and it brings 2K resolution with optimised reference control for 2D comics and motion graphics

MiniMax H3 has arrived in CapCut, and its arrival is not a quiet one. The official @capcutapp announcement promises 2K high-resolution output and what CapCut is calling optimised reference control, aimed specifically at creators building 2D comics and motion graphics. That is a meaningfully targeted pitch: H3 is not being sold as an all-purpose video generator here but as a specialist tool for a genre that has historically been underserved by AI video platforms chasing photorealism.

The rollout covers every surface CapCut offers. On mobile, H3 is available through AI Lab, AI Generator and AI Video. On desktop, it sits inside AI Video. On the web, it lives in Video Studio. That breadth matters: a creator who starts a project on their phone and finishes it at a workstation does not need to switch platforms or lose their workflow context. Availability is currently limited to subscribers in Asia, the Middle East, Africa, Europe and South America, which suggests North American access is either being staged or is subject to a separate regulatory consideration.

What MiniMax H3 actually is

MiniMax H3 is an open-weights video generation model from Chinese lab MiniMax. It has been generating considerable interest in the local-inference community: professor and AI researcher @emollick ran it entirely on a local machine and produced a photorealistic clip of an otter using a laptop on a plane in roughly three minutes, calling the pace of progress "rapid." Separately, multiple creators and observers noted the model's motion graphics output was visually striking, with PixVerse's official account flagging H3 motion graphics as "a strong hook" in the intro sequences they had seen.

The headline capability for CapCut's integration is the reference control system. Most AI video tools accept a reference image as a loose stylistic suggestion. H3's reference control is described as "optimised," implying tighter adherence to source material, which is precisely what comic and motion-graphics work demands: a character must look the same in panel two as in panel one, and a brand's colour palette cannot drift across a ten-second bumper. Whether CapCut's implementation of that reference control matches what local runners report from the raw weights is an open question, and something worth testing in the first session.

How this sits against current alternatives

The 2D comics and motion-graphics use case has been picked up by only a handful of AI video tools with any seriousness. Pika has focused on stylised effects and audio. Kling and Runway have chased cinematic realism. Seedance 2.5, which dominates the current conversation, is exceptional at photorealistic human motion but is not the first tool most comic artists would reach for when they want flat, cel-style output with consistent character design. H3 inside CapCut is targeting a gap, not a crowded lane.

The 2K resolution ceiling is a practical upgrade. Much of the AI video field has only recently consolidated around 1080p as a reliable standard output. 2K gives the artist headroom to crop, reframe and scale for print-adjacent uses, which is directly relevant if the output is going into a motion comic, a graphic novel trailer or a branded short.

What is not yet known

No pricing has been announced beyond the subscription requirement. CapCut operates on a credit model and the cost per generation for H3 at 2K is not stated in the launch post. Generation length limits are also not specified, which matters for anyone planning longer-form comic sequences. The geographic restriction to non-North-American regions is unexplained, and no timeline for expansion has been given.

The "reference control" language is also deliberately vague. It is not clear whether this means image-to-video with strict identity preservation, style reference, or a multi-reference system where separate images can drive character and background independently. Testing the limits of that system early will tell you more than any spec sheet.

What to do with this

If you are in one of the available regions and already have a CapCut subscription, the path in is straightforward: AI Lab on mobile, AI Video on desktop or desktop app. Start with a single character on a plain background and a reference image you have used before in another tool. That gives you an immediate like-for-like comparison on consistency. Then try a two-character scene, because reference control systems almost always degrade when there are two subjects competing for the model's attention. Motion graphics creators should test a simple logo reveal or typographic animation to see whether the flat-design aesthetic holds at 2K without the model trying to texture or volumise the shapes.

For filmmakers who are not in the current rollout region, the open-weights nature of H3 means local runners can already access the base model via tools like WanGP or direct Hugging Face download, albeit without CapCut's interface polish or the specific reference-control tuning CapCut claims to have applied.

---

Sources: CapCut announcement, @emollick local run demo, PixVerse community reaction