How to create consistent characters with Kling O3 reference-to-video on Venice

What reference-to-video makes possible

Kling O3 reference-to-video (R2V) is now available on Venice, and it solves the problem that has limited AI video: keeping the same character, object, and scene looking identical from shot to shot. The guide builds a 30-second luxury product commercial for a fictional brass privacy key called the "chiave" to demonstrate this.

Watching the finished piece, the point is consistency. The same device shows its book, circuit tree, griffin, and orb detail at every angle. The protagonist keeps the same hair and earrings across shots, and a circuit-tree motif even reappears on her hand. With good reference material, the model treats your characters, objects, and scenes as fixed elements and carries them through an entire story.

Start with a brief and a mega prompt

Because this is a commercial, the work starts with a brief, not prompts. The example brief, "Sovereign: Your Keys, Your Kingdom," describes a woman in her mid-30s moving through a Venetian palazzo at dusk carrying the chiave, walking from a private study to a stone balcony over the lagoon, ending on a product hero shot. It is widescreen, about 30 seconds, with audio, and a four-beat arc: discovery, passage, arrival, reveal, told across five or six shots.

To turn that brief into usable prompts, Venice published a tutorial prompts repo (linked below the video) with a reference-to-video asset and prompt generator. Copy the mega prompt into a Venice chat and fill in the placeholders: concept, characters, object or product, environment, and story arc. It returns every prompt you need for the shots and the reference images. There is also an installable skill: give it a story or ad idea and it produces the arc, reference image prompts, and the rest.

Build four-angle reference images with Nano Banana 2

Consistency depends on giving Kling enough visual information. For each element, a character, object, or scene, generate four reference images: frontal, 45 degrees to the left, a side view, and a rear view. These let the model understand the subject from every angle. You also need scene reference images (the study, the balcony) and a start frame, such as the character seen from behind in her study.

The recommended model for these images is Nano Banana 2, which gave the best results in testing, though you can experiment with others. Paste the prompts from the mega prompt into Nano Banana 2 and generate them one by one. To get a missing angle, either prompt for it directly or edit an existing image:

Shift the perspective so we see the device from a 45 degree angle.

Getting clean angles can take a few tries, but this image work is what produces the consistency Kling can then carry through the video.

Tag elements in the Venice interface

In Venice, choose image-to-video and select the new Kling O3 Standard R2V model. Upload your start frame (the study scene), then begin adding elements. When you add an element, it gets tagged, for example "element one, frontal," and you then add its three remaining angle references.

Add each element the same way: the chiave as one element with its four angles, the character as another with hers. The key step is tagging the right images so the model knows which reference belongs to which element, then writing the shot prompt that references those tags. The first shot prompt pulls in scene one, the chiave, the character's hand, and the orb:

Warm candlelight, a woman's hand with face, a woman's hand,
the blue orb, subtle dust particles, no camera movement.

Send that generation off and move on to the next shot. Reuse the same element references for following shots; not every shot needs a start frame, so some can run on element references alone.

Run multi-shot generations and switch models when needed

Kling can generate a multi-shot sequence in one go. The example sets up shots A, B, and C at five seconds each, adds the balcony scene reference, and tags the elements so the character (element two) and the chiave (element one) stay consistent. The prompt drives the action across the shots, for example element one stepping through the archway onto the balcony with the lagoon behind her and wind in her hair.

For the final product-orbit shot, the guide switches to VEO 3.1 instead of Kling, because Kling does not handle on-screen text well. Using the frontal chiave image as the reference, the prompt calls for a slow 360-degree orbit on velvet and polished marble, plus a text overlay, set to eight seconds:

A slow 360 degree orbit around the chiave brass device,
standing upright on velvet and polished marble.
Text overlay: Your keys, your kingdom.

Add music and finish in post

With the shots done, generate a soundtrack from the same brief. In a new Venice chat, click the plus, choose generate audio, and paste the music prompt. Several models are available; the guide uses Eleven Labs Music, set to about one minute, instrument-only.

The final result took many generations and regenerations of video, music, and images. The shots were assembled and trimmed in an external editor (Final Cut Pro), a title was added at the end, and the music was reworked to fit the mood, with only one shot keeping its originally generated audio. Reference-to-video does the heavy lifting on consistency, but a finished piece still depends on cutting and arranging the clips yourself. Running these generations requires credits in your Venice account.

Key takeaways

• Kling O3 reference-to-video keeps characters, objects, and scenes consistent across shots when you give it strong reference images.
• Start from a brief and use the Venice mega prompt or skill to turn it into every shot and image prompt you need.
• Generate four angles (frontal, 45 degrees, side, rear) per element, ideally with Nano Banana 2, plus scene references and a start frame.
• Tag elements correctly in the Venice interface, switch to VEO 3.1 for text overlays, add an Eleven Labs soundtrack, and finish the cut in a video editor.