Boreal-H3 Reference to Video

Boreal-H3 Reference to Video — Reference-Guided AI Video Generation

What is Boreal-H3 Reference to Video?

Boreal-H3 Reference to Video is Creatify's reference-guided video model, a post-train of the MiniMax H3 base tuned for advertising and short-form social video. You supply up to nine reference images, three reference videos and three reference audio tracks, name them in the prompt as Image 1, Video 1 and Audio 1, and the model returns a finished clip that keeps the same creator, product and setting from the first frame to the last. Its signature is the look: real-camera, phone-footage realism — vlog, UGC-ad and documentary — rather than a graded cinematic sheen, with native dialogue and sound on every clip.

Key Features

• •Reference consistency from up to 9 images, 3 videos and 3 audio files, named Image 1 / Video 1 / Audio 1.
• •Keeps a character's face, a product and a room recognizable across the whole clip.
• •Native dialogue and ambient sound on every render — no separate audio pass.
• •Output at 768p, 1088p or 2k (1440 px short edge), 24 fps, H.264 video with AAC audio.
• •Whole-second duration from 1 to 15, and six aspect ratios including vertical 9:16.
• •Deterministic: identical requests return identical clips (no seed); prompt_enhancement can rewrite your prompt or keep it literal.

Best Use Cases

Reference-to-video suits UGC-style ads and social posts where the same person or product must stay on brand: talking-head creators, product demos, founder and testimonial clips, and repeatable brand characters. In testing, one image of a shop owner, one of a product and one of the room produced a handheld vertical clip in which all three carried over faithfully while she spoke to camera.

Prompt Tips and Output Quality

Write the prompt as a shot: subject, action, camera, a quoted line for dialogue, and named sounds for the audio bed. Reference every asset explicitly as Image 1, Video 1 or Audio 1. Keep prompt_enhancement on auto for a polished rewrite, or none to keep wording literal and reproducible. Avoid prompts that name real brands, logos or real people. Audio returns loud.

How to Use Boreal-H3 Reference to Video

Boreal-H3 Reference to Video turns a set of references into a finished, talking clip that keeps your creator, product and setting consistent. You give it images, optional short videos and optional audio, then describe the shot in the prompt — naming each reference so the model knows what to hold on screen.

Writing the prompt

Write the prompt like a shot list: who or what is in frame, the action, the camera, and any spoken line in quotes. Name every reference exactly — Image 1, Image 2, Video 1, Audio 1 — so identity, product and setting map to the right asset. Name the ambient sounds you want and they populate the audio bed.

Choosing references

Use images for identity, a product and a room; a short reference video when motion, camera or framing should follow a specific clip; and reference audio to carry a voice or sound. Clear, well-lit images give the strongest consistency. Reference audio needs at least one image or video alongside it.

Recommended settings

• •resolution: 768p to iterate quickly; 1088p or 2k for finished clips.
• •duration: whole seconds, 1 to 15; 5 is a solid default. A 15-second request returns about 14.4 seconds.
• •aspect_ratio: 9:16 for social and UGC; auto follows the first reference image; set it explicitly for a 1088p clip that uses only video references.
• •prompt_enhancement: auto for a polished rewrite (adds latency, non-literal); none to keep wording exact and reproducible.
• •No seed — identical requests return identical clips, so vary the prompt to explore.

Tips

The default example used three images — a creator, a product and a studio — and all three carried into the moving, talking clip. Put a quoted line in the prompt for native dialogue, and expect loud audio. Avoid naming real brands, logos or real people.

Examples to try

Each example ran end to end at 768p, 5 seconds, 9:16; what you name in the prompt carries into the clip.

Reference-guided examples

Name every reference by type and order — Image 1, Video 1, Audio 1 — so the model knows what to hold; you can use up to 9 images, 3 videos and 3 audio tracks.

• •Identity, product and setting from three images (default). Prompt: "Handheld phone video of the woman from Image 1 at her wrapping table in the flower studio from Image 3, building the bouquet from Image 2 - coral ranunculus, white anemones and eucalyptus. She snips a stem, tucks it in, then glances up at the camera and says, "The coral is what pulls it all together." Daylight from the shopfront window, shears snipping and paper rustling." All three — florist, bouquet and studio — carry over, and she speaks to camera with native sound. Tip: give one clear image each for person, product and room.
• •A reference video drives the camera. Prompt: "Slow handheld phone clip circling the retro rotary telephone from Image 1 on a sunlit wooden desk; the camera arc and gentle sway follow Video 1, moving around the rotary dial and the coiled cord before settling. Quiet room tone and the soft click of the dial." The phone holds while the camera takes its arc and sway from the clip, not the prompt. Tip: add a reference video only when a specific camera move matters, and keep it short.
• •Product and setting with cause-and-effect, no dialogue. Prompt: "A spoon taps and cracks the glassy torched-sugar crust of the creme brulee from Image 1, set on the marble cafe table from Image 2; close handheld phone framing, shards splinter and the custard shows underneath. Daylight across the marble; the sharp crack of caramel and quiet cafe room tone, no voices." The ramekin and marble stay put as the crust cracks and a shard lifts, the crack on the native audio. Tip: skip the speaker — name the action and the sounds.

Also worth trying: "the presenter from Image 1 holds up the product from Image 2 and talks to camera in a bright kitchen".

Does it generate sound?

Yes — native dialogue and ambient sound are on every clip. Name the line and sounds you want in the prompt.

How many references can I use?

Up to 9 images, 3 videos and 3 audio tracks, referenced as Image 1, Video 1 and Audio 1.

Can it keep my product consistent?

Yes — give a clear product image and reference it in the prompt; the product stays recognizable throughout.

Is there a seed?

No. Identical requests return identical clips, so change the prompt for a different result.

What resolutions and lengths are supported?

768p, 1088p or 2k, at any whole-second length from 1 to 15 (15 seconds returns about 14.4).

How is it different from the siblings?

Text-to-video is prompt-only; image-to-video animates one start frame; reference-to-video builds the shot from multiple images, videos and audio.