Create MiniMax H3 Video Online

MiniMax H3 is a multimodal video model released as Hailuo 03, and MiniMax Hailuo H3 refers to the same model. It reads text, images, video, and audio as a single context, then generates one 4–15 second shot with native synchronized audio at 768P or 2K. Videoify runs it as a hosted service through an upstream API provider, so you work in the browser with no model files, no environment to configure, and no GPU of your own.

Open the generator at the top of this page, choose text-to-video, image-to-video, or reference-to-video, then write a prompt and add any required media. Pick the duration, resolution, and aspect ratio, and the credit cost appears before you submit. There is no setup and nothing to install. You can compose a prompt while signed out and the draft is restored after you sign in, but generating needs an account and enough credits. New accounts receive 10 signup credits. Invite bonuses and daily check-in credits are coming later. MiniMax H3 costs 2 credits per second at 768P or 3 credits per second at 2K.

Describe one shot in this order: subject, action, setting, camera movement, lighting, style, and ending state. Say what should change during the clip instead of restating what a reference image already shows. Prompts accept up to 7,000 characters, but a focused paragraph usually beats a long stack of adjectives. When you attach references, name what each file is for — this image is the character, this one is the location, this video is the movement — so the model is not guessing which part to copy.

Put the spoken line in quotation marks and say who delivers it, in what tone, and when it should start — for example, the woman turns to camera and says "we open at six," calm, near the end of the shot. Keep it short, because a 4–15 second clip has room for one or two lines at most. Write the line in the language you want to hear, but treat the output as generated audio rather than a guaranteed reading: pronunciation, timing, and lip sync vary between runs, and we do not publish a supported-language list. If a line comes out garbled, shorten it, simplify the wording, name the speaker more clearly, and generate again.

Ask for the speed you want directly in the prompt. Terms like real-time, natural speed, or brisk push back against the drift toward slow motion, while cinematic, dreamy, graceful, and floating tend to invite it. Remove any slow-motion wording you did not mean literally, and describe an action with a clear beginning and end, since someone crossing a room reads faster than someone standing still. A shorter duration also helps, because MiniMax H3 has less time to stretch the movement. Pacing still varies between runs, so compare two or three takes before changing anything else.

For image-to-video, upload a start frame in the composer and optionally add an end frame. For reference-to-video, switch the workflow first, because the image, video, and audio reference slots only appear in that mode. Images accept JPG, PNG, or WebP up to 30 MB, videos accept MP4 or MOV up to 50 MB, and audio accepts MP3 or WAV up to 15 MB. You can attach more than one file per slot, but loading everything you have rarely helps: use the smallest set that clearly shows the character, wardrobe, location, or movement that matters, and say in the prompt what each file is for. Extra or conflicting references usually make the result less predictable. MiniMax H3 reference-to-video handles the conditioning on its own, so no ControlNet or other add-on is involved.

One generation produces one clip of 4 to 15 whole seconds. No setting extends a single MiniMax H3 generation past 15 seconds, and there is no free path around that limit. To build something longer, plan the story as separate shots, reuse the same reference media so the character and setting stay recognizable, generate each shot on its own, and assemble them in a video editor.

Not as editing operations. Reference-to-video accepts MP4 or MOV files, but it uses them to guide movement, timing, and camera direction in a newly generated shot rather than converting your footage frame by frame, and it will not preserve every detail of the source. There is no timeline, no inpainting brush, and no object-level repair here. Use a dedicated editor for cuts, captions, transitions, retouching, or exact frame changes.

Give the model unobstructed reference images of the same character — face, hairstyle, wardrobe, and any distinctive detail — shot from angles close to what you want on screen. Repeat those details in the prompt instead of assuming the images carry them, and reuse the strongest references across every shot in a sequence. Video references can guide movement and audio can guide timing. This is how MiniMax H3 character consistency improves in practice, but it stays guidance rather than a lock: a full character replacement will not land identically in every frame, so generate a few takes and keep the one that holds.

Two layers apply. Videoify's acceptable use terms rule out illegal activity, infringement of intellectual-property, privacy, or publicity rights, sexual content involving minors, non-consensual intimate content, impersonation, harassment, and misleading synthetic media released without a disclosure the law requires. Separately, the upstream provider running the model applies its own filters, which can reject a prompt or a reference file that our terms would technically allow — most often around real public figures, graphic violence, and explicit material. A failed generation returns your credits automatically. We do not publish a filter list and the boundary can shift without notice, so when MiniMax H3 refuses a prompt, rewrite the sensitive part rather than trying to route around the filter.